Refactor: LensDataset and WrapperDataset for Cleaner Data Handling

[Apprentice2907]

Closes #123

What was changed

• Removed duplicate imports in dataset.py
• Extracted repeated download/check logic into _resolve_dataset_dir() helper
• Replaced hardcoded if label == 0: image = image[0] with configurable npy_index_map
• Added WrapperDataset for dynamic transform injection at runtime
• Unified DeepLenseDatasetSSL under LensDatasetSSL extending LensDataset
• Added DefaultDatasetSetup (non-SSL) class
• Added full type hints throughout
• Renamed utils/random.py → utils/lens_random.py to fix Windows circular import bug
• Updated all imports across main.py, main_ray.py, finetune_byol.py,
  simsiam.py, utils/inference.py, self_supervised/examples/byol/load_pretrained_byol.py

Testing

• Verified imports resolve correctly
• Tested with .npy simulation data, 224x224 tensor outputs confirmed