Add focal loss option for classification

[jmduarte]

• Add focal loss option for classification

Note I explicitly keep the current implementation F.cross_entropy if self.options.classification_focal_gamma == 0 out of an abundance of caution even though I checked that, numerically, the more complex formula gives the same answer.

The difference is that F.cross_entropy fuses F.log_softmax and F.nll_loss, so it is supposed to be more numerically stable, which I wanted to keep. But if you think it's unnecessary, we can remove this if-then clause.

@mstamenk

[jmduarte]

@Alexanders101 is this PR ok for you? Can you merge it? Or do you have any requested changes?