Approximated Behavioral Metric-based State Projection for Federated Reinforcement Learning