refactor(transcription): introduce AsrEngine trait (Track-B foundation)

[farce1]

What — Track-B step 1 of N

The foundation for pluggable ASR engines (the path to Parakeet-TDT-v3). A behavior-preserving refactor that puts the existing sherpa Whisper recognizer behind a small trait:

pub trait AsrEngine {
    fn transcribe(&mut self, sample_rate: u32, samples: &[f32]) -> AsrOutput; // { text, language }
}

• New transcription/engine.rs: AsrEngine, AsrOutput, WhisperEngine (wraps WhisperRecognizer).
• worker.rs now drives a Box<dyn AsrEngine> (loses ~14 lines; engine construction moved into engine.rs). The VAD → chunk → transcribe loop was already engine-agnostic.
• Behavior-preserving: identical WhisperConfig (lang auto, 2 threads, CPU) and identical text/language mapping.

A new engine (Parakeet) now plugs in by implementing AsrEngine + a loader — no further worker.rs surgery.

Verification (TDD)

• engine.rs test: trait is object-safe and returns engine output (via a FakeEngine, which also demonstrates the Parakeet extension point).
• clippy --all-targets --all-features -- -D warnings ✅; full cargo test → only the pre-existing model_archive_consts_tests (REPLACE_WITH_ placeholders) fail.

Track-B roadmap (this is step 1)

1. AsrEngine trait ← this PR
2. ParakeetEngine via sherpa-rs OfflineRecognizer (transducer) — Parakeet-TDT-v3
3. Model registry + download for Parakeet
4. Engine-selection setting + plumbing (default Parakeet-v3, Whisper-turbo fallback)
5. Model-picker UI
6. GPU execution providers (separate, larger)
7. retranscribe_meeting over stored OGG

Merge note

worker.rs overlaps with the Track-A PRs #4 (segment construction), #6 (resampler/sample-rate), #8 (panic-guard wraps the transcribe call). Integration: the panic-guard should wrap engine.transcribe, and build_segment_result consumes AsrOutput. Suggest merging Track-A first, then rebasing this.

⚠️ CI red on maintainer-only blockers (vendor binary + REPLACE_WITH_ placeholders), not this PR. Code-only, no new deps (Parakeet will reuse sherpa-rs).