Update evaluation and inference code to handle other precisions and models

[coryMosaicML]

Currently inference and eval code assumes amp_fp16 at inference which is no longer correct for some models. This PR makes this somewhat configurable, and fixes some precision bugs during inference associated with the the t5 text encoder.

Changes:

• Update clean_fid_eval.py and evaluate.py to run using the image dataset, which allows the same dataset config to be used regardless of the model type. Note this is a config breaking change.
• Add support for multiple remotes in the image dataset
• Add option to specify precision in generate_geneval_images.py
• Add a dtype option to the generic ModelInference class
• Disable autocast in precomputed_text_latent_diffusion.py for the computation of text embeddings at inference, and also fix a minor sequence length bug.