Refactor Classification Model_I and Model_II dataset loading for portability and deterministic class mapping

[Panchadip-128]

Summary

This PR refactors the dataset loading logic in the Classification pipeline (Model_I and Model_II) to improve portability, reproducibility, and contributor usability.

Motivation

Previously:

• Dataset paths depended on user-specific directory structures.
• glob usage was loosely defined.
• Class index mapping relied on dictionary iteration order.
• No validation existed for empty dataset directories.

These issues reduced portability and made onboarding difficult for new contributors.

Changes

• Standardized dataset structure to:
  root_dir/class_name/*.npy

• Replaced ambiguous glob usage with:
  os.path.join(root_dir, "", ".npy")

• Added validation to raise a clear error if no .npy files are found.

• Made class mapping deterministic using sorted class names to ensure consistent label indices across runs.

• Improved path handling using os.path utilities for cross-platform compatibility.

Impact

• No changes to model architecture
• No changes to training logic
• No changes to transform behavior
• No changes to output format

This is purely a structural and usability improvement.

Expected Dataset Structure

Example:

data/
Model_I/
axion/
cdm/
no_sub/
Model_I_test/
axion/
cdm/
no_sub/