Implement a proper gliner client, fix bugs

[Ingvarstep]

Summary

Rewrites GLiNERClient as a pure HTTP client, drops the Ray dependency from the client path, fixes a torch.compile-related packing bug in the span-relex model, cleans up GLiNERFactory.shutdown(), and substantially expands the serving docs (RelEx section + client usage).

Details

gliner/serve/client.py — HTTP-only client

• Replaces the ray / serve.get_deployment_handle based client with a stdlib-only HTTP wrapper built on urllib.request.
• New GLiNERClientError exception type surfaces HTTP and network failures.
• New constructor signature: GLiNERClient(base_url, route_prefix, timeout, max_concurrency) — no longer needs a Ray cluster or deployment handle.
• predict(...) and predict_async(...):
  • str in → dict out; list in → list out (preserved).
  • Batch calls fan out one HTTP request per text concurrently (threads for sync, asyncio.to_thread + asyncio.gather for async) so the server-side @serve.batch coalesces them into a single forward pass. A sequential loop would serialize on the wire and defeat batching.
  • max_concurrency bounds the in-flight request count.
• _build_payload(...) centralizes request shape; optional fields (relations, threshold, relation_threshold) are only included when set.
• get_client(...) updated to mirror the new constructor.

gliner/modeling/base.py — simpler packing in UniEncoderSpanRelexModel._pack

• Drops the torch.compiler.is_compiling() branch that previously avoided a GPU→CPU sync in eager mode at the cost of skipping packing.
• Always computes lengths = rep_mask.sum(dim=-1) and max_len = lengths.max().item(), so eager execution now benefits from packing too.

gliner/serve/server.py — cleaner shutdown

• GLiNERFactory.shutdown() now calls ray.shutdown() after serve.shutdown() when Ray is initialized.
• Prevents noisy ServeController ... killed by ray.kill retry warnings in the raylet log on process exit. Still idempotent.

docs/serving.md

• Expands the GLiNERClient section: clarifies it’s a pure HTTP client (no Ray import), documents the custom constructor args, and explains the batching semantics with an example.
• Adds a full Relation Extraction section covering RelEx server startup (--model knowledgator/gliner-relex-large-v1.0), client usage (sync + batched), in-process GLiNERFactory usage, a curl example, and the response shape.
• Documents new/changed CLI defaults: --batch-wait-timeout-ms 50 → 10, --target-memory-fraction 0.8 → 0.9, plus --memory-overhead-factor 1.3.

.gitignore

• Adds labels_trie.cpp (generated artifact) to the ignore list.

Notable behavior changes

• Client no longer needs Ray. Any previous code constructing GLiNERClient(deployment_name=..., ray_address=...) must migrate to GLiNERClient(base_url=..., route_prefix=...).
• Eager mode packs length-variable inputs in UniEncoderSpanRelexModel, incurring a GPU→CPU sync but reducing downstream compute on masked positions.
• GLiNERFactory context manager exit now fully tears down Ray, not just Serve.