Fast-FedUL: A Training-Free Federated Unlearning with Provable Skew Resilience