Feature request: Multi-GPU dataset mapping for SDXL training

### Feature request

We need to speed up SDXL dataset pre-process. Please make it possible to use multiple GPUs for the [official SDXL trainer](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_sdxl.py) :)

### Motivation

Pre-computing 3 million of images takes around 2 days.
Would be nice to be able to be able to do multi-GPU (or even better – multi-GPU + multi-node) vae and embedding precompute...

### Your contribution

I'm not sure i can wrap my head around the multi-GPU mapping...
Plus it's too expensive for me to take x2 A100 and spend a day just figuring out the staff since I don't have a job right now.