limit crop_to_factor

[abred]

To avoid stitching artifacts it is sufficient to only crop_to_factor on the last/highest level.
It is not necessary to crop during training.
However, during training the output size has to be strictly larger than crop_factor (prod(downsample_factors))
(I don't have a pytorch setup at hand right now)

[abred]

ok, cropping only on the last level does not seem to be the best way, but cropping at every level is neither (as the final output size is then smaller than necessary in a number of cases)
One way seems to be to compute, at the bottleneck, the 'naive' max output size (without cropping) and the max 'correct' output size (largest multiple of prod(downsample_factors) smaller or equal). And then check at each upsampling level, if the difference between those two divided by product of remaining downsample_factors is >= 1. If yes, crop that and update the difference value (diff = diff - (diff//prod)*prod)