feat: add OpenAI-compatible POST /v1/audio/transcriptions endpoint

[Kartvya69]

Closes #140

What

Exposes POST /v1/audio/transcriptions by wrapping Gemini's audio input capability into an OpenAI-compatible transcription API, so tools hardcoded to the OpenAI endpoint (e.g. OpenWhispr) work unmodified.

How

• Minimal hand-rolled multipart/form-data parser (RFC 7578, binary-safe) — no new dependencies
• Extracts the uploaded audio file and converts it to Gemini inlineData
• Maps OpenAI request fields where possible:
  • model → Gemini model (suffix parsing reused)
  • prompt → guiding context for the transcription
  • language → language hint
  • response_format → output encoding
  • temperature noted as out of scope for prompted transcription
• Custom transcription system prompt instructing the model to return only the transcription text
• Response formats:
  • json (default) → {"text": "..."}
  • text → plain text
  • verbose_json → OpenAI verbose schema (duration, language, segments[] with start/end/text/...) — segment placeholder fields (tokens, avg_logprob, ...) are zeroed since prompted transcription cannot produce them
  • srt / vtt → valid subtitle output
• Timestamps: for verbose_json/srt/vtt an extra instruction asks the model to emit [MM:SS.mmm -> MM:SS.mmm] text lines, which are parsed into segments. If the model omits timestamps, falls back to a single segment spanning the audio
• Audio duration estimation (MP3 frame walk incl. VBR + ID3v2 skip, WAV header, Ogg EOS page) for fallback end times — noted in-code as a ceiling; swap for a real probe if exact durations matter
• Reuses the existing retry / account-switch / request-tracking pipeline

Out of scope (as discussed in #140)

Streaming transcription, diarization. Timestamps come from model prompting, not a decoder — good but not sample-exact.

Testing

Tested end-to-end against a live instance (v1.3.5) with gemini-3.7-flash on a short sound-effect MP3 (~1.8s):

• json → {"text":"Aaaaah!"}
• text → Aaaaah!
• verbose_json → {"duration":1.776,"segments":[{"start":0,"end":1.813,"text":"Fah!"}],...}
• srt / vtt → valid timestamped subtitle output
• Missing file field → 400 invalid_request_error

Transcribed wording varies between runs on ambiguous sound-effect audio (model interpretation), which matches expected prompted-transcription behavior.