Distributed Quasi-Newton Method for Fair and Fast Federated Learning