feat: add in-process OpenAI-compatible STT API service

[krystophny]

Summary

This PR adds an in-process OpenAI-compatible STT HTTP service to Voxtype. The service runs alongside the daemon and reuses the daemon's transcriber, so Voxtype does not load the Whisper model twice when both push-to-talk STT and an API endpoint are needed.

Scope

• Adds /healthz, /v1/audio/transcriptions, and /v1/audio/translations.
• Adds service config, CLI flags, and environment overrides for bind host/port, request timeout, upload limits, and allowed languages.
• Adds request-level language and prompt overrides for the Whisper transcriber.
• Adds json, text, and verbose_json response handling, including segment timestamps for long-form chunking.
• Keeps a bounded Whisper state pool for concurrent service requests.

This branch is intentionally independent from macOS support. It is based on main and does not include the macOS files from #129. For local Mac testing with both feature sets stacked, use feature/macos-openai-stt-stack.

Design Notes

• No auth is implemented in the local Voxtype service; the intended deployment is loopback/private LAN first.
• The service can constrain accepted request languages through config.
• Audio upload decode/downmix/resample to 16 kHz mono is handled server-side.

Verification

Executed on mailuefterl in /tmp/voxtype-openai.J5NRzM/repo after applying this branch on top of main:

$ cargo check
Finished `dev` profile

$ cargo test
549 unit tests passed; 25 integration tests passed; 0 failed

$ cargo build
Finished `dev` profile

Stacked Mac branch verification on feature/macos-openai-stt-stack:

$ cargo build --release --features gpu-metal
Finished `release` profile

$ curl -fsS http://127.0.0.1:8427/healthz
{"status":"ok"}

$ curl -fsS -F file=@tests/fixtures/vad/speech_hello.wav -F model=large-v3-turbo -F response_format=json -F language=en http://127.0.0.1:8427/v1/audio/transcriptions
{"text":"Hello world"}

Closes #244