port prototype transforms kernel tests to new architecture

[pmeier]

Per title. There are still some references to port, but I'll do that in a follow up PR. Currently there are also quite a few failing tests. I'll have a look if I botched the sample inputs generation or if the test actually uncovered some bugs.

[pmeier]

I'll have a look if I botched the sample inputs generation or if the test actually uncovered some bugs.

Both actually.

1. Some kernels don't support more than one batch dimension and I messed that up when porting the sample_inputs_fn. Still, our kernels should support arbitrary batch shapes, so I left # FIXME comments. Here is an overview of those that don't

   • perspective_{image_tensor, mask}
   • elastic_{image_tensor, mask}
   • gaussian_blur_image_tensor
   • equalize_image_tensor
   • adjust_sharpness_image_tensor

2. Some PIL kernels don't support sequences of a single element for fill if the number of channels is larger. The tensor counterpart interprets this the same as if the number were passed outside a sequence. Not sure if that is a bug and thus I didn't open an issue yet. I left some # FIXME comments. Offenders:

   • pad_image_pil
   • perspective_image_pil

3. Behavior of PIL and tensor F.pad actually diverges. Since we only wrap the stable kernels here, this is not a prototype bug. See F.crop behavior diverges between tensor and PIL images #6613.

[vfdev-5]

Looks good to me.
Let's enable crop_image_tensor and it is good to go

[vfdev-5]

@pmeier can we port gaussian blur to new framework such that we may have less flaky issues with this test :)