Predict a binary value for each transition

[mrdrozdov]

Notable differences:

• Use sigmoid rather than log_softmax in many cases. This is because BCELoss applies the log for you.
• To get reduce probabilities, we do something like (t_probs - t_preds).abs().
• To determine shift/reduce given probability, we do something like (1 - t_probs).round(). This is because probs represent the probability of shifting. This should probably be adjusted to simply round, aka should be probability of reducing.
• Entropy takes one additional step because it's not sufficient to sum like we used to.

[mrdrozdov]

The supervised model gets 83/93 on class/transition accuracy after 65k steps (so roughly the same as before). I think this should be okay to merge, unless we want to run some RL specific sanity check.

[mrdrozdov]

This is probably too outdated. Closing.