Wasserstein Distance add to loss function

[muellerdo]

Try out combination of cross-entropy/focal, dice/tversky, distance measure like wasserstein.

Run extensive experiments on various datasets. Compare loss performances.

Interesting for AutoML