feat: integrate pyannote-cloud into listener2 plugin

Summary

Adds pyannote-cloud as a new batch transcription provider in the listener2 plugin. Unlike existing providers that use WebSocket streaming via owhisper-client adapters, pyannote uses a REST-based async job model:

1. Upload local audio file to pyannote's temporary storage via media:// presigned URLs
2. Submit a diarization+transcription job (POST /v1/diarize with transcription: true)
3. Poll GET /v1/jobs/{jobId} until completion (2s interval, 10min timeout)
4. Map pyannote's TranscriptionSegment results into the existing owhisper_interface::batch::Response format

Changes:

• plugins/listener2/Cargo.toml — added hypr-pyannote-cloud, reqwest, serde_json, uuid deps
• plugins/listener2/src/error.rs — added Pyannote(String) error variant
• plugins/listener2/src/ext.rs — added BatchProvider::Pyannote variant + ~250 lines of implementation (make_pyannote_client, pyannote_upload_audio, pyannote_poll_job, pyannote_diarization_to_batch_response, run_batch_pyannote)

Review & Testing Checklist for Human

• Response mapping correctness: pyannote_diarization_to_batch_response prefers word_level_transcription over turn_level_transcription. Verify that the word-level segments from pyannote actually have per-word granularity (not full sentences), and that the space-joined transcript makes sense for both levels.
• Confidence is hardcoded to 1.0: pyannote doesn't provide per-word confidence. Verify downstream consumers handle this synthetic value correctly and don't misinterpret it.
• listen_params / model are not forwarded: run_batch_pyannote ignores languages, keywords, and model from BatchParams. The DiarizeRequest.model is set to None (defaults to precision-2). Decide if params.model should be forwarded.
• Entire audio file read into memory (tokio::fs::read): Could be an issue for very large recordings. Other providers use chunked streaming.
• Test end-to-end with a real pyannote API key: This code compiles but was not runtime-tested. The upload → diarize → poll → map flow needs real-world verification.

Notes

• The GetJobByIdResponse is #[serde(untagged)] with DiarizationJob as the first variant, which means it will greedily match diarization responses. This works for the diarize use case but the enum ordering matters.
• The pyannote client is built with a custom reqwest::Client that includes Authorization: Bearer {api_key} headers. The presigned URL upload uses a separate unauthenticated client (correct behavior).
• Poll timeout is 10 minutes (PYANNOTE_POLL_TIMEOUT), poll interval is 2 seconds (PYANNOTE_POLL_INTERVAL).

Link to Devin run: https://app.devin.ai/sessions/855b2a5608234efaab84b8aaec1a9550
Requested by: @yujonglee