Accelerating Heterogeneous Federated Learning with Closed-form Classifiers