perf: Cache Whisper model context between transcriptions

## What
Avoid reloading the selected Whisper model for every transcription command.

## Why
- `.planning/codebase/CONCERNS.md` notes that `transcribe_audio` creates a fresh `WhisperContext` for each run.
- Loading large models repeatedly adds latency and disk I/O.
- Model lifetime is currently scoped to a single command call.

## How
- Introduce a controlled model cache keyed by Whisper model size.
- Keep cache ownership behind a worker or state object that accounts for memory pressure.
- Evict or reload when the selected model changes.
- Add tests for model-cache selection logic without requiring real model files.