[bug] scorer_utils

[vince62s]

There are two issues with the scorer (used at validation time):

1. here: https://github.com/OpenNMT/OpenNMT-py/blob/master/onmt/utils/scoring_utils.py#L49-L58
   we remove all EOS/BOS/PAD tokens from the batch to reconstruct a "genuine" batch that will be passed to the translator.
   However we shouls BOS/EOS the way it was added (depending on the task MT/LM, and on the side SRC/TGT)
   Indeed, when there is such a token in the prefix or suffix strings, it is removed as it should not.
   Exemple: nllb-200, it requires a suffix like " deu_Latn"

2. when translating requires "translation options" (egs: tgt_file_prefix, but might but true for some others) they are not taken into account since we are in "training mode" wrt the "opts" cases in opts.py
   This may require some adaptation.

@l-k-11235