feat: add --diarize_device flag to allow separate device for diarization

[tioback]

Problem

ctranslate2 (faster-whisper) does not support MPS, which forces --device cpu for the entire pipeline on Apple Silicon. However, pyannote.audio (diarization) is pure PyTorch and does support MPS.

The result is that diarization runs on CPU unnecessarily, making it the dominant bottleneck. This affects all Apple Silicon users and is the root cause of the performance issues reported in #109 and #1283.

More generally, the transcription and diarization backends have different device support matrices — there is no reason to force them onto the same device.

Benchmark on Apple Silicon M-series, 5-min audio:

Step	CPU	MPS
Diarization	~300s	~21s

Solution

Add a --diarize_device flag that, when specified, overrides --device only for DiarizationPipeline. Defaults to None, which falls back to --device — fully backward compatible.

Changes

• whisperx/__main__.py — new --diarize_device argument in the diarization params group
• whisperx/transcribe.py — pop diarize_device and pass it to DiarizationPipeline
• tests/test_diarize_device.py — 8 tests covering CLI parsing and device routing (all mocked, no model downloads required)

Usage

# Apple Silicon: transcription on CPU (required by ctranslate2), diarization on MPS
whisperx audio.m4a --device cpu --diarize_device mps --diarize

# Multi-GPU: transcription on GPU 0, diarization on GPU 1
whisperx audio.m4a --device cuda --diarize_device cuda:1 --diarize

# Default behavior unchanged
whisperx audio.m4a --device cuda --diarize

Closes #109

[tioback]

@m-bain - can you take a look at this PR, please?

[Barabazs]

Great contribution, thanks!

Please drop the tests this PR. It's not needed for this type of change.

[Barabazs]

Suggested change

	logger.info(f"Using model: {diarize_model_name}")
	logger.info(f"Using model: {diarize_model_name}")
	logger.info(f"Diarization device: {diarize_device}")

Adding a log statement would be helpful.

[Barabazs]

Suggested change

	parser.add_argument("--diarize_device", default=None, type=str, help="Device to use for diarization, overriding --device (e.g. 'mps' on Apple Silicon). Useful because ctranslate2 (transcription) does not support MPS but pyannote (diarization) does. Defaults to --device if not specified.")
	parser.add_argument("--diarize_device", default=None, type=str, help="Device to use for diarization (e.g. cuda, mps, ...)")

Please keep the help string concise.