Prebuilt: repin the PR set on Inkling and Kimi-K3

[danielhanchen]

Fixes the nightly. Run 30401043818 failed in Resolve tag and every build job was skipped:

CONFLICT (content): Merge conflict in src/models/models.h
CONFLICT (content): Merge conflict in tests/test-backend-ops.cpp
unslothai/llama.cpp#40 (233cedbe) does not merge cleanly onto b10165 + the PRs listed before it

#40 was a combined branch carrying MiniMax-M3 and Inkling. Upstream master now ships its own MiniMax-M3 (ggml-org#24908), so #40's copy collides with it. ggml-org#24523 is closed, so that entry was already inert.

New set

#	PR	pin
1	ggml-org#24423 DiffusionGemma	c3fb9724 (unchanged)
2	ggml-org#25731 TML Inkling	79ca732e
3	#44 Kimi-K3 fixes + MoonViT-3d	23fac110
4	ggml-org#26185 Kimi-K3 text	cf67f0d2

Both ggml-org#25731 and #44 were mergeable: false before this; their conflicts are fixed on their own branches rather than in a fork-side mirror, so the pins point at the real PRs.

• Add TML Inkling architecture ggml-org/llama.cpp#25731 was cut before MiniMax-M3, Nanbeige4.2, the WebGPU arch-skip rework and the RPC v5 bump landed. All conflicts were add/add collisions at the same insertion point, kept on both sides. The one judgement call is RPC_PROTO_PATCH_VERSION: master reset it to 0 when it went to proto 5 at 101 ops, Inkling adds an op, so it is now 1.
• kimi-k3: full-size model fixes and the MoonViT-3d vision tower #44 sits on model: add Kimi-K3 text model ggml-org/llama.cpp#26185's head, which predates the Parakeet clip graph and the WebGPU skip-list rework. For tests/test-llama-archs.cpp, upstream replaced the skip list with DEEPSEEK32/GLM_DSA after fixing the binding alias issues (ggml-webgpu: Fix some binding alias issues to support all archs, fix recurrent-state-rollback test ggml-org/llama.cpp#25931); KIMI_K3 is appended to that list since upstream never evaluated it.
• kimi-k3: full-size model fixes and the MoonViT-3d vision tower #44 also merges Add TML Inkling architecture ggml-org/llama.cpp#25731. Inkling and Kimi-K3 both register a new arch in llm_arch_is_hybrid, llm_arch_supports_sm_tensor, llama_model_mapping, llama_model_rope_type and tools/mtmd/CMakeLists.txt, so they cannot go in as two independent heads.

Ordering

ggml-org#26185 has to come after #44, not before it. #44 was rebased onto ggml-org#26185's head, so ggml-org#26185 is an ancestor and the resolver reports Already up to date. Listed before #44 it conflicts on tests/test-llama-archs.cpp. It stays in the set so the release manifest names it, but note it is a strict no-op for the build and it is the one entry that can break the nightly for nothing if pwilkin force-pushes the pin away.

Verification

Replayed the resolver's merge loop locally against b10173 (the tag latest resolves to right now), in list order:

--- ggml-org#24423   ok
--- ggml-org#25731   ok
--- unslothai#44     ok
--- ggml-org#26185   Already up to date.
FINAL unmerged=[]

The merged tree builds clean: cmake -B build -DGGML_CUDA=OFF -DLLAMA_CURL=OFF -DLLAMA_BUILD_TESTS=ON && cmake --build build -j 28, exit 0. The lint checks in unsloth-pr-set-lint.yml were replayed locally and pass with no errors and no warnings: all four are open, all four pins are commits of their PR, all four pins equal their head.

Supersedes #43, which dropped Inkling along with MiniMax. #40 is no longer pinned.