Load latent file without original image

Hi,

I'm using a cloud environment to do training.
To minimize the upload file size, I only upload the latent files and the meta_lat.json to the environment, but the current dataset handler does not allow this.

I make the following change to train_util.py:47, and training works fine.
`IMAGE_EXTENSIONS = [".png", ".jpg", ".jpeg", ".webp", ".bmp", ".npz"]`

From my understanding, this change should not break anything.
Please consider this use case in your next update, thx.

Start by inspecting train_util.py around line 47, where IMAGE_EXTENSIONS is defined, and review how the dataset handler recognizes latent files. Confirm that a dataset containing only latent files and meta_lat.json can be used for training without affecting image-based datasets; no test file is mentioned in the issue.

Written by the indexing model from the issue text.