Make special (but common) case of k=2 more efficient

[bhancock8]

For k > 2, it's good to have our task head output k logits that can be softmaxed. But when k=2, it slows us down and hurts stability to be calculating two logits instead of just one. When k=2, I think we can have the head output size be 1 but then in predictions inflate it to a 2D tensor so that it stays compatible with the rest of MeTaL. It may be as simple as two if statements, or we may decide to subclass EndModel into BinaryEndModel.

[vincentschen]

The age-old question! For now, the Snorkel Classification is the latest implementation of the classifier package. Feel free to re-raise an issue in that repo if still relevant!