Introduce config-driven dataset and dataloader modules for Vision Transformer.

[rashi-ops]

This PR introduces a modular, config-driven data pipeline for the Vision Transformer experiments by separating dataset and dataloader logic into reusable components.

Issues:

• Hard-coded dataset paths and notebook-specific data handling.
• Dataset and dataloader logic tightly coupled with training code.

Changes:

• Added 'dataset/my_dataset_vit' to handle .npy gravitational lensing data.
• Added 'dataloaders/dataloaders.py' with 'create_dataloaders(cfg)' function.
• Added Config-driven paths, batch size, and transforms in 'lensid.yaml'.

Previously, dataset and dataloader logic were hard-coded inside the notebook.
This is making reuse and experimentation difficult.
This PR decouples data handling from training, improves reusability.