Support OpenAI stop sequences in server

[eloe]

Summary

Fixes #1044.

This adds OpenAI-compatible stop handling to the MLX-VLM server for /chat/completions, plus matching mlx-vlm compatibility support for /responses so both server endpoints honor the same caller intent.

The implementation normalizes stop as either a string or a list of one to four non-empty strings, then incrementally filters decoded text so stop sequences are trimmed even when they span token/chunk boundaries. It is wired through continuous batching, speculative decoding, and non-batching fallback paths. When a server-side stop sequence is matched, generation is cancelled/removed from the active batch and the response reports finish_reason: "stop". When a requested stop sequence is not emitted before generation exhausts, chat completions report finish_reason: "length".

I also fixed a related Responses streaming consistency issue discovered while testing: response.completed.response.output_text now matches the final trimmed response.output_text.done text.

OpenAI spec validation

Validated against the current official OpenAI documented OpenAPI spec at https://app.stainless.com/api/spec/documented/openai/openapi.documented.yml.

• Chat Completions defines stop via StopConfiguration: nullable string or array with minItems: 1 and maxItems: 4; returned text must not contain the stop sequence.
• Chat completion finish_reason includes stop for a natural stop or provided stop sequence, and length when the token limit is reached.
• Streaming chat examples finish with a final chunk whose delta is empty and whose finish_reason is stop.
• The current OpenAI CreateResponse schema does not define a top-level stop parameter, so /responses handling here is intentionally treated as mlx-vlm compatibility behavior rather than a claim that OpenAI Responses currently accepts top-level stop.

Validation

Automated tests:

uv run --with pytest python -m pytest mlx_vlm/tests/test_server.py
# 59 passed, 3 warnings

Additional checks:

python3 -m compileall -q mlx_vlm/server.py mlx_vlm/tests/test_server.py
git diff --check

Live worktree integration validation against mlx-community/Qwen3.6-35B-A3B-nvfp4 on the local MLX-VLM server:

• Chat non-streaming string stop: passed
• Chat streaming string stop: passed
• Responses non-streaming string stop: passed
• Responses streaming string stop: passed
• Chat newline stop: passed
• Responses newline stop: passed
• No matching stop reaches length: passed
• List stop chooses earliest emitted stop: passed
• Rejects more than four stop strings: passed
• Rejects empty stop string: passed

Integration sweep result: 10 passed, 0 failed.

Broader repo note: pytest -q still fails during collection on an unrelated existing mismatch in mlx_vlm/tests/test_utils.py, which imports get_class_predicate from mlx_vlm.utils although that symbol is not present.