feat(transcribe): add MPS device support for Apple Silicon GPU acceleration

[missabawse]

Context

The current hardware detection (scripts/config/hardware.ts) maps all macOS machines to mlDevice: 'cpu' since CUDA is unavailable on Apple Silicon. As a result, transcribe-audio.js runs faster-whisper on CPU on M2/M3 Macs.

Apple Silicon GPUs expose Metal Performance Shaders (MPS) which can accelerate PyTorch workloads. The original OpenAI Whisper library supports --device mps. However, faster-whisper (CTranslate2) does not currently support MPS — see openai/whisper#2504.

Options to investigate

1. Switch to openai-whisper on Apple Silicon — uses PyTorch MPS backend, supports --device mps. Trade-off: slower than faster-whisper on CUDA, no built-in VAD filter.
2. Switch to mlx-whisper — Apple's MLX framework, natively targets Apple Silicon GPU (both CPU+GPU unified memory). Fastest option on M2/M3. Requires a new Python script.
3. Wait for CTranslate2 MPS support — track CTranslate2 issue tracker for MPS backend addition.

Proposed change

• Add supportsMps: boolean and update mlDevice to 'cuda' | 'mps' | 'cpu' in HardwareProfile
• Map darwin + arm64 → mlDevice: 'mps' once a supported backend is chosen
• Add a new transcribe-mps.py script (mlx-whisper) or update transcribe-gpu.py to handle --device mps

Acceptance criteria

• detectHardware() returns mlDevice: 'mps' on darwin/arm64
• Transcription uses Apple Silicon GPU on M2/M3
• VAD filter equivalent is available on the chosen backend
• Unit tests updated for the new mlDevice values

Reference

• openai/whisper discussion: How do I use my GPU instead of CPU for Whisper on an M2 Mac? openai/whisper#2504
• MLX Whisper: https://github.com/ml-explore/mlx-examples/tree/main/whisper