Implement mapData adversary

[ngoodman]

GANs are interesting, and viewed as a learning algorithm, they are applicable to some webppl programs with model params. As a first step, just playing around with the adversarial classifier could be neat.

For the cases where the target function contains only observes (as in dream learning), we can define an adversarial classifier that tries to tell apart model forward samples from data samples. (The objective is the usual classifier cross entropy.) The average of the objective for this classifier gives us a different (from elbo) measure of how well our model is capturing the data distribution.

Note: GANs don't usually have anything to do with guide / recognition models, but i wonder if the adversary and the guide could be combined in some interesting ways?