feat: add MLX Whisper backend for Apple Silicon

[ben0724-ACE]

Summary

Add an optional MLX Whisper backend for efficient local transcription on Apple Silicon Macs.

• Support --whisper-mode mlx
• Add mlx-whisper as an optional [mlx] dependency
• Map existing model names to MLX Community checkpoints
• Preserve the existing SRT, Markdown, and video-cutting workflow
• Add MLX-specific tests and documentation

Compatibility

This backend is opt-in. The default whisper mode and the existing Whisper, Faster Whisper, and OpenAI backends remain unchanged.

MLX mode passes media files directly to mlx-whisper and skips the PyTorch/Silero VAD path.

Testing

Tested on an Apple Silicon Mac with Python 3.12.13 and mlx-whisper 0.4.3.

• black --check .
• pytest test/test_mlx.py test/test_cut.py — 8 passed
• Successfully transcribed an MP4 using mlx-community/whisper-large-v3-turbo
• Verified mixed Chinese and English SRT/Markdown output