allow training with augmentation

[jlehnen]

Some transormations like RandomCrop return PIL Images rather than Tensors, that our current framework does not allow. Probably not hard or time consuming but needs to be done

[weidler]

Was indeed rather trivial. The ToTensor and Normalize transforms need to be in the end! Then the dataloader first augments, then converts to tensor and can return the tensor instead of a PIL image.