fix: bound realtime websocket state for long sessions

[LauraGPT]

Summary

• retain only a rolling audio window while keeping absolute session timestamps
• cap speaker chunk/embedding history and preserve stable speaker IDs with bounded centroids
• move blocking session work off the asyncio loop while serializing access to shared models
• release raw audio on finalization, including sub-chunk sessions

Fixes #3101

Validation

• python3 -m pytest tests/test_realtime_ws_service.py -q (13 passed)
• real funasr/SciPy/sklearn speaker clustering smoke test (8 chunks, stable speaker 0)
• synthetic 2-hour, 16 kHz stream: 72,000 frames in 1.640s; retained buffer 80,000 samples / 320,000 bytes; reported duration 7,200,000ms
• git diff --check
• python3 -m py_compile examples/industrial_data_pretraining/fun_asr_nano/serve_realtime_ws.py tests/test_realtime_ws_service.py

[gemini-code-assist]

Code Review

This pull request introduces memory optimizations and performance improvements to the realtime ASR service, including bounding the speaker tracker's history with a deque, implementing audio buffer compaction to prevent memory leaks, and running blocking operations off the main event loop. A correctness issue was identified in the speaker tracker's finalization logic where sentences straddling the history boundary can lead to incorrect speaker assignments; a fix is suggested to split these sentences at the boundary.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.