fix(room-io): preserve audio frame processors across stream transitions

[nightcityblade]

Fixes #5448

Summary

• stop _ParticipantInputStream._close_stream() from closing the active frame processor during ordinary stream transitions
• close replaced per-track processors inside _ParticipantAudioInputStream._create_stream() and close orphaned processors when a track disappears without replacement
• add a regression test covering a direct rtc.FrameProcessor passed via noise_cancellation so it survives the initial subscription and still closes during final teardown

Testing

• .venv/bin/python -m pytest tests/test_room_io_noise_cancellation.py tests/test_room_io_teardown.py
• .venv/bin/ruff check livekit-agents/livekit/agents/voice/room_io/_input.py tests/test_room_io_noise_cancellation.py tests/test_room_io_teardown.py

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

nightcityblade seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[nightcityblade]

Thanks, good catch. I restored the selector-created processor cleanup for participant changes that do not immediately create a replacement stream, including and switching to a participant with no available track yet. I also added regression coverage for both paths and reran the targeted room I/O noise cancellation tests.

[longcw]

thanks for the pr! If you don't mind I created a separate one with some cleanup and a different impl in #5467

[nightcityblade]

Thanks for the heads-up, and thanks for putting together the follow-up PR with the cleanup and alternate implementation. That sounds good to me. Since you've got #5467 moving forward, I'm happy to defer to that one.