Add Deepgram (nova-3) as a third transcription backend

[apicurius]

Summary

Adds Deepgram as a third option alongside the existing Groq + OpenAI Whisper backends. Useful primarily because Deepgram's /v1/listen has no per-request size limit — Whisper APIs cap at 25 MB, which limits long-video coverage even with mono 16 kHz audio.

Why nova-3 / utterances

results.utterances[] already gives us pre-segmented chunks with start, end, and transcript, mapping cleanly onto the existing {start, end, text} segment shape used by both the VTT parser (transcribe.parse_vtt) and the Whisper verbose_json adapter (_segments_from_response). No downstream changes needed in transcribe.filter_range or format_transcript.

smart_format=true&punctuate=true&detect_language=true keeps behavior parity with Whisper's defaults (auto language, punctuated output).

Wire-level differences from Whisper

The Deepgram client mirrors _post_whisper's retry/backoff envelope (4 attempts, 2 of them on 429), but differs on three points:

1. Auth header is Token <key>, not Bearer <key>.
2. Body is the raw audio bytes — no multipart form.
3. Response shape is results.utterances[] (or results.channels[0].alternatives[0].transcript as fallback).

Pure stdlib — no deepgram-sdk dependency, consistent with the existing Groq/OpenAI implementation.

Backend selection

Preference order when multiple keys are set: Groq → OpenAI → Deepgram. Override with --whisper {groq,openai,deepgram}. setup.py scaffolds DEEPGRAM_API_KEY= alongside the other placeholders and accepts it as satisfying the preflight key check.

Stderr messages and docs refer to "transcription" / "speech-to-text" rather than "Whisper" where the broader concept applies — but the --whisper CLI flag name is preserved for back-compat.

Test plan

• python3 -c "import ast; ast.parse(open('scripts/whisper.py').read())" — syntax clean
• Smoke-test against an X.com video without captions: extracted audio, uploaded to api.deepgram.com/v1/listen, got back 26 segments aligned to the speaker's actual delivery.
• Verified _segments_from_deepgram_response falls back to the alternative transcript when utterances is absent (manual response stub).
• setup.py --check returns 0 with only DEEPGRAM_API_KEY set.
• setup.py --json reports whisper_backend: "deepgram".

Notes

• Bumped plugin.json to 0.2.0 (additive feature, no breaking changes to default backend behavior).
• No new dependencies. No changes to the frames.py / download.py paths.