Add support for VibeVoice 1.5B / 7B

[samwestb]

I see VibeVoice Realtime and VibeVoice ASR are supported but the original models don’t seem to be.

https://huggingface.co/microsoft/VibeVoice-1.5B

[Talpik]

I've successfully implemented a converter to the mlx format. Where you can also get both the full and quantized versions of the model. I completely copied the 1-to-1 inference logic from the VibeVoice community repository. I get excellent results on the q6 version. However, I'm having performance issues. The full model generates 0.25x (the original generates 0.75x), while the q6 model generates 1.25x on my Mac M4Max hardware.

[Talpik]

Added Draft PR #640