fix(rtc): filter media events from the room FFI subscription

[swayamg20]

Summary

Fixes #812.

Filter audio_stream_event and video_stream_event from the long-lived subscription created by Room.connect(). Audio/video streams already have their own subscriptions, so forwarding these events to every room only adds event-loop callbacks and room-queue work.

Keep all other events eligible, including the publish_track and unpublish_track callbacks that LocalParticipant receives through the room queue. The temporary connect subscription and AudioSource.capture_frame() are unchanged.

Thanks to @wizd for the report and the correction identifying the publication-callback requirement. The separate publication-state robustness issue is intentionally out of scope. This change also leaves the connection-cancellation work in #813 untouched.

Regression coverage

The new tests exercise the real room connection, FFI queue, room listener, and participant publication methods with a fake native request boundary:

• Audio/video events schedule no room callbacks while still reaching an independent media subscriber.
• Room events, RPC invocations, publication callbacks, and audio capture callbacks still arrive.
• Public publish/unpublish calls complete, and disconnect cleans up subscriptions.

Reverting only the filter makes both media regressions fail: 100 media events schedule 200 callbacks instead of the expected 100. Restoring it passes all eight new cases. This is deterministic dispatch-count evidence, not a production throughput benchmark.

Validation

• Focused regression and existing FFI queue tests: 15 passed, with warnings treated as errors.
• Full tests/ and livekit-rtc/tests/ suite: 121 passed, 72 skipped on Python 3.13.12/macOS arm64. Live-server credentials were unset.
• Repository-wide Ruff lint and formatting: passed.
• Mypy across protocol/API/RTC packages and the new test: passed, 100 source files.
• RTC wheel build and git diff --check: passed.

Local fixture note: git-lfs is unavailable in this checkout. The full-suite run used the exact upstream WAV files in a temporary directory, verified against the tracked LFS SHA-256 hashes; a pytest collection hook redirected only the three audio test modules' fixture directory. No assertions or tracked fixtures were changed. The existing JWT test-key and AudioStream test-double destructor warnings remain (34 warnings). Skipped/live-server tests have not been validated.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[cloudwebrtc]

lgtm