On the Inference Calibration of Neural Machine Translation