Add deeplense_npy_loader.py with robust loader for inconsistent .npy formats

[abhiram123467]

Fixes #178

Problem

DeepLense .npy files exist in 3 different formats causing
ValueError for new contributors loading the dataset.

Solution

Added deeplense_npy_loader.py with:

• load_npy_image() — handles all 3 .npy formats automatically
• LensingDataset — clean PyTorch Dataset class with train/val split

Formats handled

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

Testing

Tested on all 3 classes (no_sub, cdm, axion)
with ~90,000 files on Google Colab T4 GPU ✅