Refactor Dataset Pipeline: Modular LensDataset & Transforms

[BeathovenGala]

Description for issue #126

The current implementation in [dataset/preprocessing_model_2.py] tightly couples data loading, hardcoded category logic (e.g., if file_name.startswith('axion')), and transformations. This has resulted in a fragile, duplicated, and untestable codebase.

Issues with current approach:

• Fragile: Breaks on new .npy datasets that don't match specific hardcoded filename prefixes.
• Duplicated: Normalization logic (e.g., Min-Max scaling) is copied across 3 different files, multiplying the risk of bugs (e.g., division by zero).
• Untestable: No isolation between loading and processing, making it impossible to verify transformations independently.

Fixes

This PR refactors the pipeline into modular, reusable components:

• Introduced [LensDataset] (pure loading) and [WrapperDataset] (handles categories dynamically via config, removing hardcoded logic).
• Added [get_transforms(config)] for a modular, configuration-driven augmentation pipeline.
• Added comprehensive test suites for datasets and pipelines.

Problem Demo

The Error (Existing Code):

# Fragile: Fails if files aren't named 'axion'
if file_name.startswith('axion'): 
    data_point = data_point[0]

#  Duplicated: Normalization logic repeated in 3 locations
# Risk of division by zero not handled consistently
normalized = (data - min) / (max - min)