Implement K-Fold cross-validation to train the U-Net

[MelanieLu]

At the moment, the training set, validation set and test set are fixed, and new annotated patches are added only to the training set after each active-learning iteration, leaving the validation set unchanged.

Using K-fold cross validation would allow to make use of all the samples to train the network, especially when few data samples are available.

One potential issue coming with the K-fold cross validation is the extension of the training time.