Run pyannote diarization on the GPU when CUDA is available (17.5x faster)

[Suijiku]

Closes #50.

Summary

diarizer.py loaded the pyannote pipeline but never moved it to a GPU, so speaker
diarization ran entirely on the CPU — even when the user selected CUDA and transcription
used the GPU. On long recordings the diarization step then dominates total time.

This threads the configured compute device into DiarizationWorker and moves the pipeline
to CUDA when available, falling back to CPU otherwise. Only the neural stages (segmentation,
embedding) move to the GPU; the final PLDA clustering stays CPU-bound regardless — but the
neural stages are the bulk of the cost on long audio.

Changes

• app/transcription/diarizer.py
  • _resolve_device(device) — returns "cuda" only when torch.cuda.is_available(), else "cpu".
  • _get_pipeline(hf_token, device="cpu") — moves the pipeline via pipeline.to(torch.device("cuda"))
    when resolved to CUDA; cache key is now (hf_token, resolved_device) (a CPU pipeline and a
    CUDA pipeline are distinct objects).
  • DiarizationWorker gains a device param and emits a "Running speaker diarization on GPU…" note.
• app/main_window.py — passes transcription.device from config to DiarizationWorker.
• tests/test_diarizer.py — device-selection tests (CUDA→GPU, CUDA-unavailable→CPU, CPU→CPU,
  per-device cache) with torch/pyannote mocked.

Notes

• No pre-Volta concern. Unlike CTranslate2 (see the compute-type PR), pyannote uses fp32,
  which pre-Volta GPUs handle fine — verified on the GTX 1080 below.
• Portable / non-regressive. CUDA-less machines (Intel/AMD) resolve to CPU exactly as before.
• Accuracy unchanged. CPU and GPU produce the same 4-speaker result on the test recording.

Testing

Measured on a GTX 1080 (Pascal), Windows 11, CUDA 12.6 PyTorch, pyannote.audio 4.0.7,
speaker-diarization-community-1, on a real 21.5-minute recording:

Path	Wall-clock	Speakers
CPU	13.32 min (799 s)	4
GPU	0.76 min (46 s)	4
Speedup	17.5×	identical

End-to-end through DiarizationWorker(device="cuda"): emits "Running speaker diarization on
GPU…", completes in ~52 s (incl. load), assigns SPEAKER_00–03, no errors.

• python -m pytest tests/ -v — full suite green (includes the new device tests).