Allow masking padding tokens in cross attention layers

[jazcollins]

This PR adds a new parameter to the stable_diffusion_xl and stable_diffusion_2 classes called mask_pad_tokens that allows for masking out padding tokens in cross attention layers.

The generate() had to get a bit more complicated due to the setting where we pass in pre-tokenized inputs.. we'd now want to allow passing the padding mask with it (as well as for pre-tokenized negative prompts). Let me know if you think of a better way of handling this :/

One small note: this change might be slightly redundant with the zero_out_negative_prompt arg (in the generate() function) and zero_dropped_captions (in dataloader) that I added not too long ago for zero-ing out empty negative prompts and dropped captions. I think mask_pad_tokens ought to serve a similar purpose of masking out the empty text embeddings in the cross attention layers, however zero_out_negative_prompt/zero_dropped_captions additionally zero-out the pooled text embedding (used in SDXL as microconditioning). I think we still want to keep that functionality. The cleanest way would prob be to merge these all into one flag?

[coryMosaicML]

LGTM, thanks for this!