Remove unnecessary input image padding

[BVlad917]

If an image's height/width already is a multiple of the window size, then we don't have to pad the input image. E.g., for an image of shape 256x512 and using window size of 8x8 the previous code would pad in both height and width by a full window in both dimensions (i.e., 8 pixels in height and 8 pixels in width), even though this is unnecessary. This is now removed, and for the given example the padding would be 0 in both dimensions.

[xenova]

This is also something I noticed today while adding this to 🤗 Transformers.js.

When following this demo, the image (which is already 256x256) is unnecessarily padded to 264x264, meaning the upscaled version also has the padding (see right and bottom, the image is mirrored):

[xenova]

Also, when running it with the replicate link (shown in the README), it does not add the padding:

[wade0604]

I tried to replicate SWIN2SR, using the same experimental configuration as SWINIR, but it seems to be less effective than SWINIR from the logs, my training set DF2k and test set is Set5
The first log below is swinir, the second is swin2sr

train.log
train (1).log