[Enhancement] Add sequence-level distillation to NMT training

[sxjscience]

Description

Add the sequence-level distillation to NMT training. This means, we draw samples from the teacher model with beam-search and train the student model with the generated samples.

References

• https://arxiv.org/pdf/2006.10369.pdf
• https://arxiv.org/abs/1606.07947