TF32 support request.

https://developer.nvidia.com/blog/accelerating-ai-training-with-tf32-tensor-cores/
https://huggingface.co/docs/transformers/v4.15.0/performance#performance-and-scalability-how-to-fit-a-bigger-model-and-train-it-faster

TF32 is a better bf16 and it would be nice to train with a higher precision.

It seems to be [trivial ](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py#L501)to add but I failed trying to add it. Because I saw no differences.

Start by reading the linked NVIDIA and Transformers references, then inspect the referenced examples/text_to_image/train_text_to_image_lora.py location and compare its behavior with this project's training options. Establish a reproducible way to verify whether TF32 is enabled and affects training; done should include a demonstrated result or a documented explanation of why no difference appears.

Written by the indexing model from the issue text.