Refactor: Split COCO dataset into detection and segmentation datasets

[Chandraveersingh1717]

Summary

Split coco_detection_dataset() into two specialized datasets for object detection and instance segmentation tasks.

Problem

• Current implementation carries 500MB+ annotation object in memory for entire dataset lifetime
• Confusing documentation mixing detection and segmentation use cases
• Tasks never run together (different model architectures)
• Poor cache organization with unidentified large files

Solution

coco_detection_dataset() - Object Detection Only

• Returns: boxes, labels, area, iscrowd
• Memory: ~250MB (50% reduction)
• Use: Faster R-CNN, YOLO, SSD

coco_segmentation_dataset() - Instance Segmentation (NEW)

• Returns: boxes, labels, area, iscrowd, segmentation, masks
• Memory: ~250MB
• Use: Mask R-CNN, DeepLab

Cache Organization: Files now stored in /coco subdirectory for better identification

Breaking Change

Segmentation users must migrate:

# Before
coco_detection_dataset(..., target_transform = target_transform_coco_masks)

# After  
coco_segmentation_dataset(..., target_transform = target_transform_coco_masks)