Long-Short Term Masking Transformer: A Simple but Effective Baseline for Document-level Neural Machine Translation