Other ANC loss functions

[hmc-cs-mdrissi]

Currently the ANC does correctness and confidence based upon take the square difference (L2). Given that the values are meant to correspond to probability distributions a different loss like cross entrop may help in training. For this to be doable the mask used must either choose to keep or ignore a row and can't choose to ignore a piece of a row (which wouldn't make sense anyways for a normal problem). Then compute the sum of the cross entropies for the relevant rows instead of the mask * (predicted - expected)^2 for those two losses.