test(inference): pin the streaming request-slot lifecycle of the steer endpoints

[Alexander230]

Summary

Rescoped after #237. The request-lock change this PR originally carried is superseded upstream: with_request_lock admits the handler, and the SSE generators take their own non-exclusive slot via stream_lock for the stream's lifetime. What was still missing is a test that pins that lifecycle, so this PR is now tests only — no production code changes.

apps/inference/tests/unit/test_steer_stream_slot.py drives /steer/completion and /steer/completion-chat with a stub eager backend (no model, no GPU) and checks that:

• the handler's admission slot is released before the stream starts, and the generator's own acquire yields the first frame promptly — an admission slot held across the stream would deadlock the single mutex a non-vLLM pod runs with;
• the slot is held while streaming and released when the stream is exhausted and when it is closed early (client gone), on both endpoints' generators;
• a non-streaming request generates under the handler's slot and takes no second one.

Verified to bite: with stream_lock mutated to never take a slot, 5 of the 6 cases fail; with the non-stream path made to take a stream slot, the deadlock is reported as a TimeoutError rather than a hang. ruff and pyright clean.

Refs #38 (closed as not planned).

🤖 Generated with Claude Code

[vercel]

@Alexander230 is attempting to deploy a commit to the Neuronpedia Team on Vercel.

A member of the Team first needs to authorize it.

[Alexander230]

CI status note, to save reviewer time triaging red checks — none of the failures on this PR are introduced by it:

• Vercel — outside-PR deploy-authorization gate; this PR doesn't touch the webapp.
• Inference - Tests: the workflow-repair commit in this PR (originally split out as Make inference CI runnable: drop the impossible 3.10 matrix leg, scope pytest to tests/ #215, consolidated here) removes the impossible 3.10 matrix leg and scopes pytest to tests/, so once approved the checks execute for real. The first still-red steps are pre-existing app-wide findings, byte-identical to main in the same environment (see table).
• Codecov / inference-coverage fails on two pre-existing tests/unit/test_sae_manager.py cases (MockSAE.to() missing 'dtype' — sae-lens mock drift), red on main since February. This PR's tests pass in that job (tests/unit/test_request_lock.py ....).

Verification from a fork rehearsal of the exact workflow (Alexander230#1, runs 29406654972 / 29406654966, Python 3.11 + 3.12 on clean runners):

CI step (app-wide)	main base	this PR	delta
ruff check .	2935	2934	−1
ruff format --check .	212 files	211	−1
pyright .	3723	3723	0
pytest tests	2 unit + 5 integration failures	same failures	+4 new passing tests

(The 5 integration failures are tokenizer/numeric drift, identical on main in the same environment.)