new functionality to capture prediction scores from the OpenNMT model

[t-shoemaker]

This PR adds a new argument to normalize_words() to optionally return scores for every candidate word outputted by the model. These scores are outputted by OpenNMT's Translator, so the changes only involve capturing this output.

Summary of changes

• Prediction scores are stored in the fake_stream class during a translation run and are called from that class after the translations are complete
• Output of call_onmt() is now a list of candidate tuples, where the first position in a tuple is the candidate word and the second is the prediction score
• Small modification to _dict_filter() involves handling the candidate tuples
• _normalize() now ends with an if/else check to determine whether the candidate words and their scores should be returned, or just the words

[mikahama]

I will merge this for now, but I still need to check the compatibility with my other libraries (UralicNLP and murre) before making a new release