【ICLR 2023】Generalization Bounds for Federated Learning: Fast Rates, Unparticipating Clients