Action Dependency Graphs for Globally Optimal Coordinated Reinforcement Learning