fix: serialize local ASR model access

[LauraGPT]

Summary

• guard the lazy SenseVoice model initialization so concurrent first requests build one shared model
• serialize AutoModel.generate() calls on that shared model
• add deterministic concurrency regressions for both behaviors

Root cause

Each WebSocket handler sends blocking ASR work to the default thread pool, while
all handlers share the module-level _sensevoice_model. The previous GIL-based
assumption did not protect the check-then-initialize sequence, so concurrent
first requests could construct multiple models.

The shared FunASR AutoModel is also stateful during generate(): it resets and
updates runtime configuration dictionaries for the model and VAD pipeline.
Allowing two executor threads to enter the same instance concurrently can race
those mutations and create unnecessary GPU memory pressure.

Both locks are acquired inside executor workers, so the aiohttp event loop
remains non-blocking. Concurrent requests queue only at the shared model
boundary.

Validation

• python -m unittest discover -s tests -p "test_*.py" -v
• python -m compileall -q server/asr_server.py tests/test_asr_server.py
• git diff --check