Implement command-line configurable data augmentations

This MR introduces command-line configurable data augmentations for training.

By default, no augmentations are added to any training session.

We also provide the following combinations of augmentations you may try:

• elastic: A simple in-package implementation of an elastic transformation with 80% probability
• affine: A set of geometrical (affine) transformations (rotation by 10 degrees, translation by 10%, scale between 80 and 100%)
• hflip: Horizontal flipping with 50% probability
• jitter: A set of jitter transforms changing brightness, contrast and saturation by 20% at most (works for both grayscale and colour images)
• hflip-affine: Combines both horizontal flip and affine transforms above
• hflip-jitter-affine: Combines horizontal flip, jitter, and affine transforms above

These changes should help us assess #67 (closed) in a more controlled manner.