Add token-level inter-annotator agreement

[kasnerz]

According to the original plan, token-level inter-annotator agreement should have been already included in this pull request:
#83

However, it is non-trivial to implement it: we need to have token-level flags of presence / absence of each error category, which needs to be pre-computed first.

We can follow the computations here, we just need to make it a bit more cleaner and efficient.