Add robust dataset loader utility for inconsistent .npy formats

[abhiram123467]

Problem

DeepLense .npy files exist in 3 different formats with no
documentation or utility to handle them:

1. Plain (64,64) float arrays
2. Object arrays of 64 row-vectors each of shape (64,)
3. Flat (4096,) arrays with label appended as extra row

This causes ValueError for new contributors trying to
load the dataset.

Solution

Add a dataset_utils.py with a robust loader that
automatically detects and handles all 3 formats.

I would like to work on this fix.