📚 The doc issue

In many geometric transformations (e.g.: ElasticTransform, RandomAffine, ...) there's a fill parameter specified, and its description states:

Pixel fill value used when the padding_mode is constant [...]

Issue is, the padding_mode parameter does not exist in the majority of these transforms' constructors! It's a very minor nitpick, but still 😉

Suggest a potential alternative/fix

The quickest fix is obviously removing the part that says 'when the padding_mode is constant', but it could also be turned into a very fun feature request!

A way to implement this functionality would be to just add a quick computation to find out the image size necessary to cover all the frame, perform the Pad transform, and finally complete the geometric transform. There might be a few edge cases though, so that could possibly require some testing.