Add TP-SDS (and potentially VSD) to improve quality/diversity

[yankeesong]

TP-SDS
This paper shows that using monotonically non-increasing time-sampling in diffusion models shows better results than uniform sampling. I can try to implement it.

[thuliu-yt16]

Of course, PR is welcome!

[yuanzhi-zhu]

@yankeesong I tried it with 2d experiments here https://github.com/yuanzhi-zhu/prolific_dreamer2d and found that Dreamtime only helps sds a bit but not vsd.

Would like to see your implementation!!

[yankeesong]

@yuanzhi-zhu I suspect this is because for SDS the diffusion is frozen whereas for VSD a LoRA is trained. If one trains the LoRA with TP (i.e. with non-uniform time sampling), they might introduce inductive bias.