Implement Pop-Art

[Learning functions across many orders of magnitudes](http://arxiv.org/abs/1602.07714) introduces Preserving Outputs Precisely, while Adaptively Rescaling Targets (Pop-Art). In summary it normalises outputs across orders of magnitudes and gets rid of the clipping (i.e. counting) rewards heuristic for Atari games. The normalisation is also better for non-stationary problems, i.e., any decent real world problem.

The below is a picture of extra notes from the authors, next to their poster at NIPS 2016:

The issue names no files, tests, or entry points. Start with the linked Pop-Art paper and the authors' poster notes, then inspect the Atari repository to locate the learning implementation. Done means Pop-Art output normalisation and adaptive rescaling are implemented without relying on reward clipping.

Written by the indexing model from the issue text.