Add mlx-whisper engine for Apple Silicon GPU acceleration

[gworley3]

Summary

• Adds a new mlx-whisper engine option that runs Whisper models on Apple Silicon GPU/Neural Engine via the mlx-whisper package
• On Apple Silicon, this provides whisper large-v3 accuracy at dramatically faster speeds than CPU-based faster-whisper (~725 frames/s warm vs ~270 frames/s on first run)
• Follows the existing engine pattern (new engine class, config field, factory entry)

Changes

• src/claude_stt/engines/mlx_whisper_engine.py — new engine class using mlx_whisper.transcribe()
• src/claude_stt/engine_factory.py — register the new engine
• src/claude_stt/config.py — add mlx-whisper as a valid engine option with mlx_whisper_model config field (defaults to mlx-community/whisper-large-v3-mlx)

Usage

[claude-stt]
engine = "mlx-whisper"
mlx_whisper_model = "mlx-community/whisper-large-v3-mlx"

Requires pip install mlx-whisper. Only works on Apple Silicon Macs.

Test plan

• Tested locally on MacBook Pro with Apple Silicon
• Verified transcription works with large-v3 MLX model
• Confirmed existing moonshine and whisper engines still work
• CI/linting (if applicable)

🤖 Generated with Claude Code