FedH2L: A Federated Learning Approach with Model and Statistical Heterogeneity