Heterogeneous Information-Bottleneck Coordination Graphs for Multi-Agent Reinforcement Learning