Created by: Superjomn

According the design of sequence decoder, a beam search operator is needed.

It takes two inputs:

• topk_ids: candidate id set,
• topk_generation_scores: the scores of the prefixes with the candidates appended. and it has two outputs:
• selected_ids: the selected candidate id set,
• selected_translation_scores: the corresponding scores of the selected prefixes with candidates appended.