`HuggingFaceNmtEngine` produces incorrect token scores when beam search is used

[ddaspit]

In order to compute the correct token scores, the method compute_transition_scores from Huggingface Transformers should be used. See the implementation in silnlp for more details.

[Enkidu93]

I'm just going to go ahead and do this as part of adding the confidence scores to pretranslations in machine.py for better overall consistency with silnlp and it's a matter of just copy-pasting a couple dozen lines of changes.