FedAR: Addressing Client Unavailability in Federated Learning with Local Update Approximation and Rectification