feat: add local transcription fallback when captions are unavailable

[gcomneno]

Summary

Implements the local-first fallback for video sources without usable captions.

Behavior

• source captions remain preferred;
• otherwise GYTE acquires only private audio with yt-dlp and transcribes it locally with Whisper;
• stable private audio/transcript artifacts support safe retry and reuse;
• pipeline provenance distinguishes source-caption from local-transcription;
• local failures are explicit and do not advance prepare;
• tests cover download/transcription, reuse, fallback selection, provenance and failure semantics.

Whisper can be configured with GYTE_WHISPER_COMMAND and GYTE_WHISPER_MODEL.

Closes #14.