BERT, mBERT, BiBERT? A Study on Contextualized Embeddings for Neural Machine Translation