Keep track of defected slices

[abailoni]

Hi @constantinpape,

I cleaned up my dev branch. These are the current additions/modifications wrt master:

• I added the possibility to keep track of artifacts added during data augmentation. I used this option to make sure that the model is trained to predict labels from the previous slice. This prevents noisy predictions, especially when a slice is completely blacked out
• I also added the options min_distance_between_defects and nb_contiguous_artifacts to the defect augmentation
• The shift option in RandomSlide can now be applied only to the raw image (leaving the GT in place)
• Minor bugfix: avoid to apply twice the normalization to artifacts (only apply to the raw image once the artifacts have been applied)