Test every nth episode

[mctigger]

Currently the experiment runner first runs the training, then the testing. However, often it is interesting to get intermediate test results while training. E.g. run (and log) a test episode every nth training episode.

This should be pretty easy to implement by adding a new Experiment class that just calls test() every nth episode in train()

[cpnota]

While this is a useful thing to have, easy to implement, and quite a few papers try to use this, there are a few issues I haven't fully thought through. For example, think about the parallel agents: if they get interrupted in the middle of the episode, how should they be resumed after some test episodes are run? I think the answer is that the test environments perhaps should actually be separate objects from the train environments, so they can maintain separate state. I'll leave this open for now!