Fix GroupBy logic in Fast Image Processor.

[yaswanth19]

@yonigozlan What do you think of the logic change. I was working on another where I had (num_crops, 3,h,w) and the existing logic was breaking coz h,w are same but num_crops are diff. This one should be no different in most cases considering num channels is first dim.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: vitmatte

[yaswanth19]

@yonigozlan for review. Here _cast_tensor_to_float func was unused hence removed it

[yonigozlan]

I don't think there's a need to change anything here. The case you described should be already supported by using is_nested=True in group_images_by_shape and reorder_images. It will flattened the batch_size and num_crops dimensions, then reconstruct them both.