CPC implementation

[ankeshanand]

CPC implementation

[sherjilozair]

This is done, but needs to be tested with probes.

I should also add frameskip in the targets, i.e. range(self.steps_to_ignore, self.steps_to_predict, self.steps_to_skip), so that longer-term predictions can be made without increasing memory much. A lot of contiguous frames are pretty redundant anyway.

Although I am using GRU for the prediction, it might make sense to concatenate 4 frames together for targets as well. In games like PacMan, there is a lot of flickering, which 4-frame-observation solves.