Brandon/cast unquantized flux to bfloat16

[brandonrising]

Summary

This PR forces unquantized flux model state dicts to be converted to bfloat16 since that's all we support for inference currently. This can be removed or minimized as we support more data types in the future. This will allow us to support the seemingly common float8 flux fine tunes available on sites like huggingface and civitAI.

QA Instructions

Attempt to perform inference with a model similar to this one

Merge Plan

Can be merged when tested and approved

Checklist

• The PR has a short but descriptive title, suitable for a changelog

[RyanJDick]

I left a handful of comments - mostly around all of the explicit memory management we are doing with (del and gc.collect()).

I tested it as-is and it works for me. The quality seems pretty clearly worse than a 4-bit quantized model though 😅