Low Precision Local Training is Enough for Federated Learning Binbin Lin