GTrans: Grouping and Fusing Transformer Layers for Neural Machine Translation