Fix: Optimize LensingDataset data loading and prevent NaN loss

[kamilansri]

Description

This PR addresses a critical mathematical flaw in the LensingDataset normalization step that could lead to division-by-zero errors (resulting in NaN loss during training). It also refactors the data loading logic to be more robust across different operating systems and significantly more memory-efficient.

Changes Made

• Fixed NaN risk in normalization: Added a small epsilon value (1e-8) to the denominator during min-max normalization. This prevents division-by-zero crashes if an image array is completely flat/blank (where min equals max).
• Robust path construction: Replaced raw string concatenation (directory+selected_class) with os.path.join to prevent malformed paths if the directory string is missing a trailing slash.
• Optimized tensor creation: Removed the redundant np.array([np.load(...)]) wrapper. Now using torch.from_numpy() to convert the array directly into a tensor without unnecessary memory copying.
• Added dimension and type safety: Appended .unsqueeze(0) to correctly add the channel dimension, and chained .float() to ensure the output tensor is float32 (saving memory and compute compared to float64).