Use COCO Mask Parsing from pycocotools

🚀 The feature

The CocoDetection v2 transform wrapper attempts to decode the mask itself, but pycocotools provides a high performance implementation already. We have had to copy from master, this _dataset_wrapper.py because of a bug related to the handling of these masks that was fixed in master but not installable using pip yet.

https://github.com/pytorch/vision/blob/main/torchvision/tv_tensors/_dataset_wrapper.py#L402

Seeing torchvision.datasets.CocoDetection has self.coco as a COCO() object, let's use it.

       coco_ann = dataset.coco.imgToAnns[image_id]

        if "masks" in target_keys:
            target["masks"] = tv_tensors.Mask(
                    torch.stack([
                        torch.from_numpy(dataset.coco.annToMask(ann))
                        for ann in coco_ann
                    ])
                )

Motivation, pitch

There have already been bugs related to this, and there's no need to reinvent the wheel. Instead, let's use the existing implementation.

Alternatives

No response

Additional context

No response

Start in torchvision/tv_tensors/_dataset_wrapper.py around the referenced mask-decoding logic and inspect the CocoDetection dataset entry point. Verify how dataset.coco.imgToAnns and annToMask provide the annotations, then run the relevant existing dataset or transform tests; done means mask targets are decoded through pycocotools without changing expected output.