Speed up preprocessing

[RossM]

Describe your changes

Pillow (PIL) supports lazy loading of images, where the actual image data is only loaded when needed. Currently image_utils.get_dim() loads the image data to handle EXIF rotations. There's no need to do that, we can just swap width and height if necessary. This gives a ~4x speedup in preload when the latents are already cached.

Issue ticket number and link (if applicable)

Checklist before requesting a review

• [ ] This is based on the /dev branch (Or a fork of it)
• [+] This was created or at least validated using a proper IDE
• [+] I have tested this code and validated any modified functions
• [N/A] I have added the appropriate documentation and hint strings if adding or changing a user-facing feature

[RossM]

Fixing a bug in the EXIF handling lost most of the performance gain. I'll look for a better way to do this.

[RossM]

I'm now loading the image size directly from the cached latents, rather than needing to open the original file. This reduces preprocessing for a particular large dataset from almost 2 hours to under 15 minutes.

This required moving loading the cached latents earlier in the process, and plumbing that through. I tried to keep as clean as I could, but the result is still pretty ugly. Please let me know if you'd like any refactoring.