fix {convert_format, clamp}_bounding_box

[pmeier]

This fixes two things that were missed in #7227 and #7228.

1. Calling a kernel should always be done by a plain tensor. Thus, we unwrap datapoints. Otherwise we are in hot water, because internally we assume that this is the case:

   vision/torchvision/prototype/transforms/functional/_meta.py

   Lines 246 to 248 in af7c6c0

   	xyxy_boxes = convert_format_bounding_box(
   	bounding_box.clone(), old_format=format, new_format=datapoints.BoundingBoxFormat.XYXY, inplace=True
   	)

   Note that the call to convert_format_bounding_box will fail if bounding_box is a datapoint.BoundingBox here, since we unconditionally pass the old_format.

2. After convert_format_bounding_box the format has to be updated and cannot be taken from the input by design

cc @vfdev-5 @bjuncek

[NicolasHug]

Thanks Philip,

Just tried this on test_detection_preset() while putting back RandomIoUCrop(), and it looks like it worked properly. Let's merge!