Bound return-audio pacing queue

[fallintoplace]

Summary

• add a per-participant return-audio backlog cap (return_audio_max_buffer_s, default 3 seconds)
• drop the oldest queued audio frames when a TTS burst exceeds the cap, keeping newer audio responsive
• expose queued/dropped frame and duration counters and rate-limit backpressure warnings
• extend pacing tests for overflow, drop-oldest behavior, flush metrics, and invalid buffer limits

Tests

• PYTHONPATH=server-runtime:utils/xr-ai-logging uv run --project agent-sdk --with pytest --with pytest-asyncio --with loguru --with numpy --with livekit --with livekit-api --with pyyaml --with fastapi --with 'uvicorn[standard]' --with httpx --with websockets --with cryptography pytest tests/test_return_audio_pacing.py\n- uvx ruff check server-runtime/xr_media_hub/transport/livekit/_room_client.py server-runtime/xr_media_hub/transport/livekit/config.py tests/test_return_audio_pacing.py