[experiment/example] Training vs. Test loss/error evaluation

[harold]

Per this interesting question from Kiran:
https://groups.google.com/forum/#!topic/clojure-cortex/8sLokBNmT2c

It would be nice to have an example of showing train vs. test loss during training to understand over/under-fitting. The current setup doesn't make it obvious how to do this, and basically everyone using the experiment training interfaces is likely to do want to do this at some point.