fix(avatar): restore the previous audio route on AvatarSession.aclose()

[dorukdumlu]

Summary

Avatar plugins route the agent session's audio to the avatar with output.replace_audio_tail(...) in start(), and aclose() never put it back. When the avatar fails to start (wait_for_join timing out, the app giving up), the session stays pointed at a dead datastream sink and the agent goes silent instead of degrading to regular audio. With the canonical ordering (avatar started before session.start()), the stale route also makes session.start() set room_options.audio_output = False, so no RoomIO audio ever exists to fall back to.

Fixes #7276

What changed

• AvatarSession gains _attach_audio_output(sink): it installs the route and remembers what replace_audio_tail() swapped out. aclose() restores it:
  • the previous tail goes back under any wrapper chain (TranscriptSynchronizer and friends stay attached),
  • a route that was installed over nothing is cleared, so a later session.start() sets up room audio normally,
  • a route someone else installed after the avatar is left alone (identity-guarded).
• AgentOutput.replace_audio_tail() now returns the sink it replaced, and AgentOutput.audio_tail exposes the current tail. Both backward compatible.
• All in-repo avatar plugins are migrated to the helper, a mechanical one-line change each. Out-of-tree plugins that call replace_audio_tail directly keep working exactly as before, just without the restore.

One known limit, called out rather than papered over: if the avatar dies after session.start() already ran, restoring the route can't bring audio back, because RoomIO audio was disabled at startup and there is no _ParticipantAudioOutput to return to. Handling that would mean building the room audio output late, which felt like a separate feature. Happy to take it as a follow-up if there's interest.

Test plan

• New tests/test_avatar_session_close.py (5 tests): restore over nothing, restore of a previous output, restore of the tail under a wrapper proxy, a newer route wins, aclose without an attach is a no-op. The restore tests fail without the fix.
• pytest tests/test_audio_sink_proxy.py tests/test_inference_avatar.py tests/test_recorder_io.py --unit (90 passed)
• ruff and mypy clean on the touched files

[dorukdumlu]

Both findings were right, addressed in acc2670. The red one was a plain miss on my side: I swept livekit-plugins for replace_audio_tail callers and forgot the one avatar that lives in core, so the inference AvatarSession kept bypassing the tracking. It now goes through _attach_audio_output like the plugins, with a close regression test on the inference path. For the yellow, the restore is synchronous and doesn't depend on the participant removal at all, so it simply runs first in aclose(), before any await; added a test that cancels the close mid-removal and checks the route still came back.