A Historical Trajectory Assisted Optimization Method for Zeroth-Order Federated Learning