Previously effectively agreement was checked on all categories,
even if a lexeme's tag didn't have some of the selected categories.
That led to treating such lexeme as not agreeing with one that had
more matching categories, even if it was agreeing on all categories
that were present.

Underlying issue was temporary table of remaining lexemes for a token not being
updated (emptied) if only underspecified lexemes of the token met an agreement.
Effectively the token was left in its original state i.e. with all
original lexemes present, both those that were underspecified and the remaining
ones that did not met a weak agreement that was otherwise met via the
underspecified lexemes.

So such lexemes are always left in the end.

Intentionally left implicit.

Rule - user-level class encapsulating WCCL rule.
Action - internal class for Rule's actions
ActionExecContext - helper class for execution of actions.

That forced antlr to run every time. The workaround is to remove the output directory
so files are generated from clean slate, thus have proper time stamps.

It is coherent now with the SentenceContext::is_inside(int abs_pos)
- SentenceContext::is_inside(Position pos) is added, while
Position::is_inside(SentenceContext) is removed

I.E. empty sentence shouldn't be passed anyway and that is what
Operator class is checking for. But otherwise, at the level of
operator, it can deal with a sentence that has current pos outside
of boundaries.