[Bugfix] Cancel inference on client disconnect + fix non-stream request leak

[yhl-amd]

Summary

Two independent bugs in the OpenAI server, both surfaced by retrying clients (opencode / Claude Code) which pile up unbounded pending requests and waste GPU on output nobody reads.

1. No disconnect cancellation. Neither the streaming generators nor the non-stream handlers aborted the engine sequence when the client hung up, so an abandoned request ran to max_tokens.

• Add an engine abort path: Sequence.aborted → Scheduler reuses the normal finish/deallocate stop path; EngineUtilityHandler.abort_request; CoreManager.abort_request (broadcast to all DP ranks).
• Streaming generators (stream_chat_response / _fanout) now run under try/finally so cleanup_streaming_request (which aborts the seq) fires on GeneratorExit.
• Starlette does not cancel the handler on disconnect for non-stream, so add _run_nonstream_with_disconnect() which races generation against raw_request.is_disconnected() and cancels on disconnect. chat_completions / completions gain a raw_request param.

2. Non-stream request leak. generate_async / _multimodal / _fanout only removed the request from io_processor.requests on a setup exception, never on normal completion → every completed non-stream request leaked a Sequence (pending grew forever). Wrap the token loops in try/finally that always pops (and aborts on early exit). Streaming already pops via cleanup_streaming_request.

Also add "choices": [] to the streaming usage chunk so AI-SDK clients (opencode) stop rejecting the final usage frame.

Why

Under an agentic client that retries / cancels mid-generation, the pre-fix server leaks a Sequence per completed non-stream request and keeps abandoned requests running to max_tokens. pending requests grows without bound and GPU stays pinned on discarded work.

Test plan

Verified on DeepSeek-V4, TP8:

• 6 sequential non-stream completions → pending requests stays flat (was 1→8).
• 3 mid-generation client disconnects → abort_request ... found=True, GPU utilization drops immediately, pending flat (was +3 leak with GPU pinned to max_tokens).
• Normal streaming + non-stream answers unchanged.

Reproduce:

PORT=9700; M=<served-model-id>
# (1) non-stream leak: fire 6, pending must stay flat
for i in $(seq 6); do
  curl -s localhost:$PORT/v1/chat/completions -H 'Content-Type: application/json' \
    -d '{"model":"'"$M"'","messages":[{"role":"user","content":"hi"}],"max_tokens":8}' >/dev/null
done
# (2) disconnect cancel: start long gen, drop after 2s -> log shows abort found, GPU drops
timeout 2 curl -s localhost:$PORT/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"'"$M"'","messages":[{"role":"user","content":"write a very long essay"}],"max_tokens":4000}' >/dev/null
# (3) regression: normal stream + non-stream still correct
curl -s localhost:$PORT/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"'"$M"'","messages":[{"role":"user","content":"2+2?"}],"max_tokens":10}'

Split out of #1428 (which now carries only the DeepSeek-V4 DSML tool-call changes) so this fix can be reviewed on its own.

🤖 Generated with Claude Code

[yhl-amd]

Split out of #1428 so the client-disconnect / non-stream-leak fix can be reviewed on its own. #1428 now carries only the DeepSeek-V4 DSML tool-call changes.