google-transcribe: numerals support via digit-sequence class token

[pragmatrix]

Why

Spoken numbers ("nine four one two") are transcribed as words ("nine four one two") instead of digits ("9412") by Google Speech-to-Text V2. Applications that need numeric output (dial pads, quantities, prices) have to post-process every transcript. Google offers speech adaptation class tokens exactly for this, and the service had no way to use them.

Summary

• Add a numerals parameter to the Google transcribe service. When set, an inline PhraseSet with the $OOV_CLASS_DIGIT_SEQUENCE class token (boost 20.0, taken over from the internal FreeSWITCH-parameterized setup) is sent as RecognitionConfig.adaptation (ADR 0005).
• Restructure google_transcribe::Params into region + flattened TranscribeParams (model, language, diarization, numerals) so recognition parameters can be passed through to the client; all new fields default, so existing configs keep working.
• Final results now take the first alternative directly: Google only populates confidence on the top alternative, so the previous confidence-based re-ranking compared unknown values. max_alternatives is hardcoded to 1 accordingly. The rejected re-ranking approach is documented in ADR 0005.
• The example CLI accepts --numerals for the Google provider.

Validation

• cargo test -p context-switch -p google-transcribe passes (4 passed, 1 ignored).
• cargo fmt -- --check clean.
• Workspace version bumped to 3.8.0 (new public service parameter).

---

AI disclosure

This pull request was developed with assistance from GitHub Copilot (GLM-5.3-Flash). All code was reviewed and approved by the human author before submission.