Optimal Federated Learning for Nonparametric Regression with Heterogeneous Distributed Differential Privacy Constraints