add proper smoke test for prototype transforms

[pmeier]

Per title. We test two things:

1. Does the structure of the output match the structure of the input?
2. If the transform can handle the input (based on Transform._transformed_types) is the type preserved? If not, does the transform pass the input through?

We test this against a plethora of configurations:

1. CPU / GPU
2. dict, tuple, list inputs
3. datapoints.Image, datapoints.Video, PIL.Image.Image, plain torch.Tensor as "core" input

Apart from the "core" input, by default each sample contains more images and videos, bounding boxes in all formats (regular and degenerate ones), detection and segmentation masks, Python builtin types (int, str), other objects (pathlib.Path), a plain tensor, and a numpy array.

cc @vfdev-5 @bjuncek

[NicolasHug]

Some comments already discussed offline

[NicolasHug]

LGTM, thanks a lot!

As discussed offline and as a follow-up we should try to validate that the degenerate bboxes stay degenerate (or potentially passed-through), although this may involve a mechanism we don't have yet