Integrate new model speech to text model Fish Speech 1.4

[jorgeantonio21]

The HuggingFace repository can be found in:

https://huggingface.co/fishaudio/fish-speech-1.4

[LaurentMazare]

Something that could be tricky here is that the license for the code is cc by nc sa, so I don't think we could have an implementation based on the original code in candle-transformers. Two possibilities there:

• Make an implementation based on the original code but have it live in another repo.
• Use as a basis another implementation which has a more permissive license or only use the research paper (not sure if any of these exist).