Fixed Encoder Self-Attention Patterns in Transformer-Based Machine Translation