feat(stt): add JSON input_audio request format

[8x22b]

What

Adds STT_REQUEST_FORMAT (multipart default | json). The json format sends base64-encoded audio in an input_audio: { data, format } JSON body instead of multipart/form-data, for STT providers whose /audio/transcriptions endpoint does not accept multipart uploads.

Why

Some providers (e.g. OpenRouter) expect audio inline as base64 JSON rather than a multipart file upload. Today this only works with a local source patch; this makes it a supported, documented option.

Changes

• config.stt.requestFormat parsed from STT_REQUEST_FORMAT (validated; falls back to multipart).
• transcribeAudio() selects the body shape based on the format. The fetch / timeout / error / response-parsing path is shared; only the request body and Content-Type differ between the two formats.
• For json, the audio format is derived from the filename extension (oga/ogg/mp3/wav/m4a/flac/aac/webm, default ogg).
• multipart behavior is unchanged.

Config

Var	Default	Notes
STT_REQUEST_FORMAT	multipart	json sends base64 audio in an input_audio JSON body

Tests

Two new cases in tests/app/services/stt-service.test.ts for the json path (body shape and Content-Type, format detection from filename, optional language). The default multipart tests are unchanged.

npm run lint, npm run build, npm test pass.