[Fix] generate: propagate the disconnect 400 instead of returning null

[AmirF194]

Motivation

Fixes #4781. The non-streaming /generate endpoint aborts the session on client disconnect but discards the 400 response it computed, so the caller gets HTTP 200 with a null body instead. Same discard-the-inner-return shape as #4776/#4777, in the sibling /generate endpoint.

Modification

generate's _inner_call() returns the disconnect error response from inside its async with block, but the call site (await _inner_call()) never captured it, and the nonlocal response it otherwise sets is only reachable on the non-disconnect path. Capture _inner_call()'s return value and return it directly when not None, falling back to response only on the normal completion path:

inner_result = await _inner_call()
if inner_result is not None:
    return inner_result
return response

Verification

Docker (python:3.11-slim, CPU-only deps, no GPU), current main (9140d372). Added test_generate_disconnect_returns_client_disconnected_response to tests/test_lmdeploy/serve/test_session_cleanup.py, mirroring the existing test_completions_v1_disconnect_returns_client_disconnected_response in the same file: a fake AsyncEngine.generate() yields one item then blocks, and a fake raw_request.is_disconnected() returns True. Confirmed the test fails on unmodified main (result is None) and passes with this change. The full tests/test_lmdeploy/serve/test_session_cleanup.py file (10 tests) stays green, and pre-commit run --files on both changed files passes.

Not verified: the CUDA/GPU CI legs (no GPU on this machine) and the pre-existing failures in tests/test_lmdeploy/serve/openai/responses/ (unrelated pydantic validation errors in the /responses API, reproduced as already red on unmodified main, untouched by this change).

BC-breaking

None.

Checklist

1. Pre-commit run on the changed files.
2. Covered by a new unit test (fails on main, passes here).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[AmirF194]

Looked at the two failing jobs (pr_functions_test, unit_test). Both fail identically before any test runs: the turbomind CUDA extension build via cmake/ninja fails during pip install -e (exit 4 and 255 respectively, same cmake --build command). This diff only touches lmdeploy/serve/openai/api_server.py and its test file, no C++/CUDA sources, so the build failure isn't coming from this change. I haven't identified the actual cause of the build failure (looks like it could be a runner-level issue, not something I can confirm from the logs alone), so I won't guess at a fix, just flagging that it predates and is independent of this diff.