Preprocessing helper for Normalized and Denormalized Boxes

🚀 Feature

We recently had box_convert function which helps interconversions of boxes.
I missed the cases where bounding boxes are normalized.
People would like to Denormalize them as well as Normalize them back.

Motivation

Motivation was same as box_convert. We need these quick helper functions which are often used through Albumentations, etc. This is boiler plate of most object detectino repos / models. Datasets come in all sort of formats.

Pitch

This can be done in 3 ways. I am not sure which one is good.

1. Have 2 seperate utilites like denormalize_boxes, normalize_boxes. Just like other box ops such as box_area, box_iou
2. Make this part of box_convert. We can have an extra argument noramalize. This seems hacky and can cause other torchscript errors. So not pretty sure if it would be nice.
3. Can this be transform? Like T.NormalizeBox() and T.DenormalizeBox() . I haven't worked with transforms closely so not sure. But can be an idea.

Alternatives

Again this is utlity. No forcing. But something useful that bounding boxes often need.

Additional context

Maybe there is a function in torchvision / workaround which I'm not aware of.

Start by locating the existing box_convert helper and related box operations in torchvision. Compare the proposed separate utilities, box_convert extension, and transforms approach, then define the API and normalization convention and add coverage for normalized and denormalized bounding boxes.