SAGDA: Achieving O(ε-2) Communication Complexity in Federated Min-Max Learning