Add batched Whisper inference with configurable batch size

[Suijiku]

Closes #46.

Summary

TalkTrack decodes sequentially. faster-whisper >= 1.1 ships BatchedInferencePipeline
(same model, VAD-chunked parallel decode), typically several times faster on long
recordings. It still returns a lazy segment generator, so the existing
for segment in segments_gen: progress.emit(...) loop keeps working (segments just
arrive in bursts) and batched segments still carry avg_logprob/start/end/text.

A new Batch size setting (Settings > Transcription, range 1–16, default 8) is
persisted in config and passed through to the worker.

Safety hatch: batch_size = 1 keeps the classic sequential path
(model.transcribe), which retains condition_on_previous_text. Any value > 1 uses
BatchedInferencePipeline. This makes the change strictly non-regressive at batch size 1.

Changes

• app/transcription/transcriber.py — batch_size on TranscriptionWorker; batched vs
  sequential path selection.
• app/ui/settings_dialog.py — "Batch size" QSpinBox (load/save).
• app/utils/config.py — transcription.batch_size default (8).
• app/main_window.py — pass batch_size from config to the worker.
• requirements.txt / pyproject.toml — faster-whisper>=1.1.0; uv.lock re-synced.
• tests/test_transcriber.py — batched vs sequential path selection, VAD enabled, segment
  mapping.

Trade-off (stated honestly)

Batching drops cross-chunk condition_on_previous_text — a minor consistency cost — in
exchange for speed and robustness against hallucination-repetition loops on long
recordings (the same trade WhisperX makes). batch_size = 1 is available for anyone who
prefers the classic behavior.

Testing

• python -m pytest tests/ -v — full suite green, including new tests that assert the
  batched pipeline is used for batch_size > 1, the sequential path for batch_size == 1,
  VAD is enabled, and segments map correctly (avg_logprob → confidence).
• uv lock --check — resolves; resolved faster-whisper version unchanged (1.2.1).
• App launches and transcribes end-to-end on the batched path (Windows 11).
• Verified against faster-whisper 1.2.1 that BatchedInferencePipeline.transcribe accepts
  batch_size/vad_filter/language and that batched segments still carry
  avg_logprob/start/end/text.

The ~4x figure is faster-whisper's documented batched-vs-sequential speedup, not a number
I measured here. Happy to add machine-specific wall-clock + peak-RAM numbers (sequential vs
batch 4 vs 8 on the same recording) if useful — just let me know the target hardware.