Inducing, Detecting and Characterising Neural Modules:

A Pipeline for Functionally Interpretable Reinforcement Learning