Weightslab Data Augmentation

[guillaume-byte]

Fix dataloader with num_workers > 0
Allow data augmentation in the data loader:

• From scratch, preloading every image and augmented images
• Live augmentation, with workers

The current problem is:

• The model will infer on augmented data, but the UI sees only original data. If affine augmentation, missmatch btw segmentation and original images..
  Solution could be:
• UI image and pred/label should pass through "dataset norm function", i.e., if defined, inverted augmentation first.

[guillaume-byte]

Current solution already allowed:

• Data Aug. initially, i.e., increasing or replacing data with augmented data.
• Live data augmentation, where basically images are stored and visualized without aug., but augmented for training