fix: create room IO media streams off the event loop and demote retryable STT errors

[muntazir868]

Closes #7441

Summary

1. Room IO: create media streams off the event loop

AudioStream.from_track / VideoStream.from_track run a synchronous FFI request, and _on_track_available called it directly from the track_subscribed handler — initializing noise cancellation (e.g. BVC()) blocked the loop for 106–119ms and tripped the loop monitor.

• _create_stream is now async; the audio/video implementations run from_track via asyncio.to_thread and pass loop= explicitly so the stream still dispatches frames onto the job loop.
• Creation runs in a task (_create_and_forward_stream) serialized by a lock, with a generation counter: a track event that supersedes an in-flight creation discards the pending stream instead of installing it, and _close_stream defers processor cleanup to that task while it settles.
• aclose waits for in-flight creations before tearing down.
• A creation failure logs and resets _track/_publication, preserving the previous synchronous failure's retry semantics.

2. Inference STT: retryable recv_task errors log at debug

A gateway reject before input end (e.g. code=2006, "Streaming ASR unavailable" during init) is retryable=True and already handled by SpeechStream's retry loop (WARNING) and the session's stream recreation. recv_task logged the same failure as ERROR twice (inline + log_exceptions), blaming the application for a handled gateway retry.

• Retryable errors now log at debug; terminal errors still log at error.
• The blanket log_exceptions wrapper on recv_task is replaced with level-aware handling.

Testing

• make check (ruff format + lint, mypy strict) — clean
• uv run pytest --unit --audio_eot — 4015 passed, 0 failed
• New tests:
  • test_audio_stream_created_off_event_loop_and_bound_to_it — creation runs on a worker thread, loop= is the running loop, and the loop keeps ticking while from_track blocks for 150ms
  • test_stream_create_discarded_when_superseded — generation guard discards an overtaken creation
  • test_retryable_error_logs_at_debug_not_error / test_non_retryable_error_still_logs_at_error — log-level behavior
• tests/test_room_io.py runs to_thread inline under virtual time (the autojump clock races real executor threads); the off-loop tests opt out with the real_time marker.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

Devin Review found 1 potential issue.

1 flag not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)