Fix QuickTalk silence gating

[cwang0810]

Summary

• Add QuickTalk PCM silence gating so fully silent chunks render idle frames without running HuBERT or the QuickTalk model.
• Add per-frame silence masking inside mixed audio chunks, so silent frames fall back to idle while voiced frames still render normally.
• Keep QuickTalk warmup able to run through the model by bypassing the silence gate for warmup only.

Validation

• pytest tests/unit/test_quicktalk_adapter.py tests/unit/test_render_pipeline.py tests/frontend/test_quicktalk_send_path.py -q
• python -m py_compile opentalking/models/quicktalk/adapter.py opentalking/pipeline/speak/render_pipeline.py tests/unit/test_quicktalk_adapter.py
• Verified on a GPU host with real quicktalk.pth and real avatars:
  • female-host-transparent pure silence: HuBERT/model calls dropped from 1/1 to 0/0, idle-frame pixel diff became 0.
  • mixed silence/voice chunk: silent mask became [true,false,true,false], only voiced frames ran model inference.
  • anime-handsome-guy pure silence: HuBERT/model calls 0/0, all output frames matched idle frames exactly.