Re-work coco data loader

A custom coco data loader was implemented in pytorch_coco.py, mostly to remove negative images from the custom COCO-like dataset.

Instead, we want to keep them as highlighted in Focal Loss for Dense Object Detection. We can also completely remove this function and use the native torchvision dataloader.

We also want to make sure that the box formats are compatible within datasets (ex. dict format for FasterRCNN), usign a custom data box transform ToFasterCNNBoxFormat.