Landan/text encoder refactor

[Landanjs]

This PR has a few new features:

1. Changes SDXLTextEncoder and SDXLTokenizer to be MultiTextEncoder and MultiTokenizer, respectively. These two new classes enable using arbitrary combinations of OpenAI CLIP, OpenCLIP, E5 and possibly other models with minor tweaks.
   a. CONFIG CHANGE: Since there is no more SDXL specific tokenizer, a new argument sdxl_conditioning is added in image_caption.py to specify whether to generate SDXL-style conditioning in the dataset.
2. Use the attention_mask in the text encoder.

Minor:

1. Set qkv_clip = None

[jazcollins]

left some comments!!

[jazcollins]

thanks for refactoring!!

since we newly have the MultiTokenizer, we will want to change this line as well to access one of the tokenizers in the list:

diffusion/diffusion/evaluation/clean_fid_eval.py

Line 163 in fa13a30

text_captions = self.tokenizer.tokenizer.batch_decode(captions[:, 0, :], skip_special_tokens=True)

[Landanjs]

Ooops, requested review before implementing .batch_decode...

[coryMosaicML]

Looks good! Thanks for doing this!

[jazcollins]

lgtm, this is awesome landan!!!

doesn't necessarily have to be in the same PR, but it would be good to update the SDXL sample yamls (both mosaic and hydra) with the new relevant yaml params (tokenizer_names_or_paths and sdxl_conditioning in the dataloader, etc.)