🚀 The feature

torchvision.ops.boxes::masks_to_boxes is used to convert a batch of binary 2D image masks to a set of bounding boxes. This proposal pertains to creating a faster and more general version of this function.

Motivation, pitch

The implementation of masks_to_boxes utilizes a for loop over the batch dimension to individually compute the bounding box for each mask, giving the function an $O(B)$ complexity and rendering it unsuitable for large batch sizes. Instead, we can recreate the function with intelligent usage of the in-build vectorized operations to achieve the same result at close to $O(1)$ complexity.

Some primitive performance benchmarking seems to validate this hypothesis.

The proposed version is also more general, as it can be extended to 3D masks easily.

Alternatives

N/A

Additional context

I've already implemented this feature and have a pull request ready. However, I wish to understand if the core contributors believe this is a worthwhile feature or not.

### Tasks