Add Mac MPS support for Gradio app

[MuyleangIng]

Summary:

• Add --device to app.py for auto/cpu/mps/cuda selection.
• Use existing resolve_runtime_device so Apple Silicon resolves to mps.
• Pass the resolved device into VoxCPM.from_pretrained.
• Disable the CUDA-only optimize path on non-CUDA devices.
• Lazy-load the ASR model only when transcript recognition is used.
• Ignore local .venv directories.

Tested:

• .venv/bin/python -m pytest tests/test_model_utils.py tests/test_cli.py tests/test_validate.py tests/test_lora_checkpoint_loading.py -q
• .venv/bin/python scripts/test_pick_runtime_dtype.py
• App probe on Apple M2 Air/macOS 15.5: device=auto resolved to mps, optimize=False, ASR/TTS models lazy before first use.