Understanding and Improving Sequence-to-Sequence Pretraining for Neural Machine Translation