add mitigation for empty bounding boxes

[pmeier]

As part of our performance optimizations we introduced torch.aminmax for bounding boxes in a few places. This works for degenerate inputs on CPU, but fails on CUDA:

>>> t = torch.rand(0, 4)
>>> torch.aminmax(t, dim=-1)
torch.return_types.aminmax(
min=tensor([]),
max=tensor([]))
>>> torch.aminmax(t.cuda(), dim=-1)
RuntimeError: iter.numel() > 0 && iter.ntensors() - iter.noutputs() == 1 && iter.noutputs() >= 1 INTERNAL ASSERT FAILED at "../aten/src/ATen/native/cuda/Reduce.cuh":1134, please report a bug to PyTorch.

This already has a report in pytorch/pytorch#85439. The consensus seems to be to also let the CPU variant raise an error. Thus, to be safe, we should simply exit early for degenerate inputs like we often do for images as well.

This error never surfaced in CI since we currently don't have CUDA CI for prototypes. I made sure it works locally.

cc @vfdev-5 @datumbox @bjuncek

[datumbox]

LGTM, thanks!