[New Model] LongCat-Video (with weights!)

### Feature Idea

LongCat-Video is a new video generation foundational model with 13.6B parameters from meituan-longcat, who have just released a LLM as well.

It differs from other models, such as Wan, in native Block Sparse Attention (practically like the recent HoloCine (add it too please :))), enabling it to generate minutes-long video without much computational cost and quality degradation.

The model is free, opensource and weights-available under MIT License.

https://huggingface.co/meituan-longcat/LongCat-Video

### Existing Solutions

https://github.com/meituan-longcat/LongCat-Video -- official repo and inference code.

### Other

Implementing Block Sparse Attention will also enable integrating the HoloCine Wan-derived model to produce Sora2-like videos with cinematic transitions. HoloCine repo: https://github.com/yihao-meng/HoloCine.