Add ViViT variant with factorized self-attention

[roydenwa]

Hi @lucidrains,

I implemented the ViViT variant with factorized self-attention (Model 3 in the paper), which learns spatio-temporal features per patch/ tube instead of the global/ frame-wise features learned in the "factorized encoder" variant. Could be useful for downstream tasks that require patch-wise features like video segmentation.

[lucidrains]

@roydenwa this looks great! thank you so much!