Faster Federated Learning with Decaying Number of Local SGD Steps