Enable batched faster-whisper transcription

[LC-86]

Summary

• forward transcribe() kwargs to the faster-whisper backend
• switch to BatchedInferencePipeline when batch_size > 1
• document the new batching path and add focused unit tests

Root cause

The current FasterWhisperModel.transcribe_core() accepts **kwargs, but never forwards them to the backend call. This means options like batch_size, beam_size, or vad_filter are silently dropped.

Because of that, callers cannot activate faster-whisper's documented batched inference path through ivrit's public transcribe() API.

What changed

• keep the existing single-file API unchanged
• lazily instantiate faster_whisper.BatchedInferencePipeline when batch_size > 1
• continue to use the plain WhisperModel path otherwise
• pass transcription kwargs through to the backend call instead of dropping them

Validation

• /Library/Frameworks/Python.framework/Versions/3.12/bin/pytest tests/test_batch_transcription.py -q
• python3 -m py_compile ivrit/audio.py tests/test_batch_transcription.py
• git diff --check

Notes

This is intentionally scoped to the faster-whisper path only. It does not try to redesign the API for multi-input batching yet; it enables the batching mechanism that faster-whisper already documents, through ivrit's existing transcribe() entrypoint.

Closes #13.

[LC-86]

Review note: I kept this intentionally narrow. The key behavioral change is that transcribe() now actually forwards batching-related kwargs into the faster-whisper path, and only switches to BatchedInferencePipeline when batch_size > 1. I also added focused tests for the previously dropped kwargs path and the batched pipeline selection. If you want, I can also add a short benchmark note in the PR description.