Induced Modularity and Community Detection for Functionally Interpretable Reinforcement Learning