Add controllable limit to size of experience replay buffer

[generic-github-user]

Training slows down over time as more state/action pairs are added.

[generic-github-user]

Better yet, simply shuffle the training data and draw a sample of size n from it when needed.