Add Block Diffusion Attention Mask to attn_gym/masks

[Skylion007]

Implement the Block Diffusion Hybrid Attention Kernel from https://arxiv.org/abs/2503.09573 . Disclaimer I am an author. It's a simple, illustrative, and efficient kernel that can achieve >5X speedup over SDPA due to the sparsity of the mask. Therefore, probably a good example to highlight flex attention for performance reasons.

[drisspg]

@Skylion007 you want to contribute it? 😊

[lapp0]

I've created a lightweight implementation of the paper and I'd be happy to contribute it (https://github.com/lapp0/nano-block-diffusion/blob/dcb77b94db33ff14500d8fdbb95d09880ed6958a/model.py#L156-L197)

Could someone advise me on which components should be included in this repo? E.g. should the construction of the noised sequence be included, or should I strictly contribute the block mask and require users to construct the required inputs themselves?