Clean up the logic behind minimum image size and padding for segmentation models

[ant0nsc]

We presently check that all images have at least a minimum image size dictated by the UNet architecture. Nevertheless, we later have padding enabled if an image is below the required patch size.
This should be cleaned up:

• If there is no padding, training on too small images should fail.
• If there is padding, training on too small images should print a warning, and then go on training (verify that this is really the case)
• Discuss what the right strategy is for testing on too small images.

AB#3904