Fixes #6670. This basically only moves the squashing that we did for the video kernels into the image kernels.

If we start to merge transforms v1 and v2, we should support this better. I haven't checked, but I guesstimate that 90% - 100% of all times we need to squash / unsquash is because we use these three ops internally:

1. vision/torchvision/transforms/functional_tensor.py

   Line 469 in 6e203b4

   img = interpolate(img, size=size, mode=interpolation, align_corners=align_corners, antialias=antialias)

2. vision/torchvision/transforms/functional_tensor.py

   Line 562 in 6e203b4

   img = grid_sample(img, grid, mode=mode, padding_mode="zeros", align_corners=False)

3. vision/torchvision/transforms/functional_tensor.py

   Line 763 in 6e203b4

   img = conv2d(img, kernel, groups=img.shape[-3])

I propose, we implement _nd_{interpolate, grid_sample, conv_2d} as internal utilities, so we don't have to have the arbitrary batch compatibility code everywhere. This is somewhat already implemented for 2. with