FedRG: Unleashing the Representation Geometry for Federated Learning with Noisy Clients