While retaining the original behavior with regards to data splits stored in .json files, I add support for nested subdirectory structures.

This works under the assumption that every point cloud MUST contain at least its coord asset (otherwise it's not a point cloud). That means that every individual point cloud's assets (normals, color, etc.) in a train/val/etc. split's directory can be located by finding the parent directory of every coord asset.

Given the following train split directory structure

train
|_pc1
|   |_coord.npy
|   |_color.npy
|_pc2
|   |_coord.npy
|   |_color.npy
|_extra
    |_pc3
        |_coord.npy
        |_color.npy

When the data split is specified as a subdirectory, and not as a json file

Old behavior: pc1 and pc2 are included in the training set, pc3 inside extra is skipped
New behavior: pc1, pc2 and pc3 are included in the training set

This is a very convenient change for constantly changing datasets, since repeatedly updating a json file is bothersome and error-prone. I've used this modification for months now without issues, so I'd like to contribute it to the community.