Fixes the incorrect token prediction distribution from _all_scores_for_token() in sequence_tagger_model.py

[mdmotaharmahtab]

This PR fixes the issue #3448 . Previously, due to incorrect length calculation for each sentence in the batch, the returned tag probability distribution for each token was incorrect (from the _all_scores_for_token() function in sequence_tagger_model.py). This PR makes a small change in _all_scores_for_token() function to correctly compute the length.

[alanakbik]

@mdmotaharmahtab thanks a lot for this PR!

@whoisjones can you review?

[mdmotaharmahtab]

I have examined the failing unit tests and found that they also failed on the master branch (so rebasing on the master did not help). They may not be related to this PR. Do I need to work on these tests for this PR? @whoisjones @alanakbik

[MdMotahar]

Hello. Want to know if this PR will be reviewed any time soon? @whoisjones @alanakbik