Proposal: Add Fish Audio (Fish TTS) provider

[iluv7]

Fish Audio (https://fish.audio) would be a valuable addition to the TTS module alongside OpenAI, Gemini, and DashScope. Its unique strengths:

• Best-in-class zero-shot voice cloning — 10–30s of reference audio reproduces a voice without training
• Multi-speaker synthesis in a single request — <|speaker:0|> tags with per-speaker reference IDs; none of the currently integrated TTS providers support this
• 83 languages with strong Chinese quality; s2.1-pro is the recommended production model and s2.1-pro-free offers a $0 dev tier
• Low latency — ~100ms time-to-first-audio and realtime WebSocket streaming
• Open weights (Fish Speech) for self-hosting

The integration fits the existing provider pattern: a FishAudioCredential, FishAudioTTSModel (REST POST /v1/tts), model cards for s2.1-pro / s2.1-pro-free / s2-pro, plus optionally a realtime model and a voice-cloning client on top of the existing VoiceCloningClientBase.

Happy to open a PR if maintainers are interested.