CI: key release wheel folders by build number instead of nixl sha (#2259

)

Release wheels published to `release/<ver>/<nixl-sha8>/` gave every
rebuild of a commit the same folder, so a re-spin overwrote its
predecessor in place and a half-finished upload was indistinguishable
from a complete one. This publishes to `release/<ver>/<build-number>/`
instead, with the nixl sha recorded as an Artifactory property rather
than in the path.

- `cuda_major` becomes a matrix axis instead of a job parameter, so one
run covers both CUDA majors (20 cells) and a build-number folder still
holds a complete cu12+cu13 set. `build-llm-container` keeps finding both
variants under a single `WHEEL_BUILD_ID`; the nightly cron drops to one
entry and `CUDA_MAJOR` is removed.
- `pipeline_stop` now owns the release folder: on SUCCESS it stamps a
folder-level `NIXL_SHA` (with `recursive=0`, so per-wheel values
survive) and points `currentBuild.description` at the folder in the
Artifactory UI; on anything else it deletes the folder, since a partial
set has to be rebuilt anyway. Safe only because the folder is keyed by
build number and owned by one build.
- Publish attaches `NIXL_SHA`, `NIXL_VERSION`, `JOB_NAME` and
`BUILD_NUMBER` to every wheel unconditionally — outside the
`build_options.env` guard, so a built ref predating
`--build-options-file` still records its sha.
- The poller's published check becomes one AQL property query per
release branch instead of one folder GET per commit, and triggers once
per commit rather than per CUDA variant. Because only a green build sets
the marker, a partial upload no longer reads as published — a gap the
old check had.
- The job description links the release tree in the Artifactory UI.
- `verification/` is untouched: those folders are keyed by sha pair and
shared between builds, so they are never deleted.

Release folders now accumulate one per re-spin and `cleanup-spec.json`
has no `release/**` retention rule. Left alone deliberately — purging
published release artifacts is a policy call rather than a CI one.

Two things need confirming against the live Artifactory on the first
run: the UI tree URL shape (`/ui/repos/tree/General/<repo>/<path>`) and
that AQL returns folder-level properties for `type: folder`.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **New Features**
- Nightly releases now build CUDA 12 and CUDA 13 wheel variants
together.
- Release artifacts are organized by version and build number for easier
identification.
- Wheels include build, source-commit, and resolved build-option
metadata.

- **Bug Fixes**
- Incomplete release folders are automatically removed after failed
builds.
- Release monitoring now detects published wheel sets more reliably and
retries missing or failed releases.
  - Partial uploads are excluded from completion checks.
  - Duplicate builds are prevented while a release is in progress.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: NirWolfer <nwolfer@nvidia.com>

test: wait-for-wheels + publish-crate on gitlab_ci_runners, PR #2076

test: wait-for-wheels DAG (nixl-compliance.yml), PR #2076

chore: update attributions for 1.4.1 (#2129)

Bump the nixl, nixl-cu12 and nixl-cu13 self-version entries from 1.4.0
to 1.4.1. No third-party dependency changes between 1.4.0 and 1.4.1:
Rust crates verified 68/68 against Cargo.lock, all Meson wrap versions
unchanged, and the wheel's torch/numpy deps are unchanged.

## What?
_Describe what this PR is doing._

## Why?
_Justification for the PR. If there is an existing issue/bug, please
reference it. For
bug fixes, the 'Why?' and 'What?' can be merged into a single item._

## How?
_It is optional, but for complex PRs, please provide information about
the design,
architecture, approach, etc._

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

nixl 1.4.0-rc3 — re-test compliance after nspect fix (MR !34)

nixl 1.4.0-rc2 — test release-automation compliance trigger

chore: 1.4.0 attributions update (#2088)

Full re-verification of the C++ dependency surface against what the
1.4.0 build actually produces (contrib/Dockerfile.manylinux wheel
builder on manylinux_2_28/AlmaLinux 8, contrib/Dockerfile container on
Ubuntu 24.04), cross-checked against the shared objects bundled in a
published nixl-cu13 wheel.

Version corrections (attributed version did not match what is
built/shipped):
- grpc 1.17.0 -> 1.73.0 (GRPC_TAG; wheel ships libgrpc++.so.1.73.0)
- protobuf 3.12.4 -> 31.0 (gRPC submodule; wheel ships
libprotobuf.so.31.0.0) and fix License URL, which pointed at Boost's
LICENSE_1_0.txt
- c-ares 1.18.1 -> 1.19.1 (gRPC v1.73.0 submodule)
- RE2 2023-03-01 -> 2022-04-01 (gRPC v1.73.0 submodule)
- LibFabric 2.3.0 -> 1.21.0 (LIBFABRIC_VERSION in all three Dockerfiles)
- libcurl 8.5.0 -> 8.19.0 (built from source for OpenSSL 3.x linkage)
- libxml2 2.9.14 -> 2.15 (built from source; wheel ships
libxml2.so.16.1.3)
- cuda 12.8 -> 13.0, cudart 12.8 -> 13.0.88, cufile 1.13.1.3 -> 1.15.1.6
- NVIDIA DOCA 3.1.0 -> 3.3.0 (DOCA_v3.3.0 host package)
- glibc 2.35 -> 2.39, libgcc-s1/libstdc++ 12.1.0 -> 14.2.0 (Ubuntu
24.04)
- libunwind 1.3.2 -> 1.3.1, Google test 1.17.0 -> 1.14.0

Missing dependencies added:
- hwloc 2.12.2 (BSD-3-Clause) - a required Meson dependency of the
libfabric plugin, built from source in the wheel builder
- openssl 3.0.16 (Apache-2.0) - built from source; gRPC, libcurl, the
AWS SDK and the Azure SDK all link it
- openldap 2.4.46 (OpenLDAP Public License v2.8) - pulled in through
libcurl

Mislabelled entry fixed:
- libucm 48.0-1 -> libibumad 48.0-1; the entry's URLs and license
already pointed at rdma-core, and libibumad was the only rdma-core
component installed by both Dockerfiles that was not yet attributed

Python: add the nixl-cu12 self-entry, which was missing while nixl and
nixl-cu13 were present, even though the cu12 wheel is built and
published.

Verified unchanged: all Rust entries (59/59 Cargo.lock registry crates),
all Meson subproject wraps, and the resolved Python dependency tree.

## What?
_Describe what this PR is doing._

## Why?
_Justification for the PR. If there is an existing issue/bug, please
reference it. For
bug fixes, the 'Why?' and 'What?' can be merged into a single item._

## How?
_It is optional, but for complex PRs, please provide information about
the design,
architecture, approach, etc._

[release/1.3.2] libfabric: flush FI_MORE per rail (#1989)

## Summary

Backports #1966 onto `release/1.3.2`.

This cherry-picks `78ece0c9897f6751052514c7cb37fde6ab7228ec` as
`9dce5cb6615e45cfd71a45fcb6f42cfa5b28dfa9`.

## Details

- Flushes FI_MORE WRITE batches per physical rail instead of by
descriptor position.
- Keeps batching opt-in via an explicit `apply_fi_more` decision passed
into `prepareAndSubmitTransfer`.
- Preserves this release branch's inline `postXfer` flow while adapting
the per-rail flush precompute from the original PR.

## Validation

- `git diff --check HEAD~1 HEAD`

Full Meson build was not run locally because `meson` / Python
`mesonbuild` is not installed in this environment.

Signed-off-by: Arun Karthik <akkart@amazon.com>
Co-authored-by: Arun Karthik <akkart@amazon.com>

chore: update NIXL attributions to 1.3.1 (#1901)

Bump the nixl and nixl-cu13 self-version headers in
ATTRIBUTIONS-Python.md from 1.3.0 to 1.3.1 to match the released package
version (pyproject is already at 1.3.1). No third-party dependency
attributions changed for the 1.3.1 patch release.

## What?
_Describe what this PR is doing._

## Why?
_Justification for the PR. If there is an existing issue/bug, please
reference it. For
bug fixes, the 'Why?' and 'What?' can be merged into a single item._

## How?
_It is optional, but for complex PRs, please provide information about
the design,
architecture, approach, etc._


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **Documentation**
* Updated license attribution entries for two packages to reflect the
latest versions.
* Kept the same Apache 2.0 license details while refreshing the recorded
package versions.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

chore: update attributions for latest release (#1783)