Federated Learning via Meta-Variational Dropout