Add mlx-whisper audio backend

[mxl]

Summary

• add an optional mlx-whisper audio transcription backend alongside the existing faster-whisper path
• make the audio backend configurable through audio.backend
• document the optional MLX installation path and backend setting

Details

• add video_analyzer/whisper_backends.py with:
  • FasterWhisperBackend
  • MlxWhisperBackend
  • create_whisper_backend(...)
• refactor AudioProcessor to use an injected backend while keeping shared extraction/transcript orchestration in one place
• add audio.backend to the packaged default config with faster-whisper as the default
• add optional mlx-whisper dependency support:
  • conditional dependency in requirements.txt
  • mlx extra in setup.py
• update docs to mention:
  • audio.backend
  • pip install \".[mlx]\"

Behavior Notes

• faster-whisper keeps the current behavior, including beam search and VAD filtering
• mlx-whisper currently uses:
  • word_timestamps=True
  • temperature=0
  • optional language
• mlx-whisper does not use beam_size because the current library raises NotImplementedError for beam search
• mlx-whisper also does not support the same VAD options as faster-whisper

Testing

• python -m pytest -q tests/test_whisper_backends.py

Notes

• A full repository-wide pytest run currently fails in the separate video-analyzer-tune subproject because dspy is not installed in this environment

[mxl]

It's better to review #63 first.

[byjlw]

It's better to review #63 first.

review comments on 63