It was returning Position based off absolute position, which
was no good for operators like orth[first(M)] working in changing
sentence contexts.

This reflects the fact that positions stored in matches are
absolute positions, while Position type is used in the
language as position relative to current position in the
sentence context. So mixing the two was confusing at the very least.
This also actually fits the most recent version of spec,
which no longer mentions Position, either.

Also, adding M in grammar to be shorthand for $m:_M (vargetter) op in match context.

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.

Previously it would mandate agreement was met on all categories
in a tag with enough categories, if such tag was found.
Now it checks for agreement only on the number of categories that
actually match, including the case of zero categories matching
(the latter does meet the weak agreement).

Clearer and better performing version, also more likely to actually work as expected :)

Note: still not implemented internals of agreement

The design doc for llook/rlook was incorrect, swapping the semantics per the updated docs and swapping argument order for llook.

Values cleanup and doxy: remove some redundancies, make the Begin/End constants in Position an enum, change tag_ref to contents

The agreement stuff isn't implemented yet.

- since the default value for Position is Nowhere now.
Also, remove reduntant header and change stream to ostream in to_strings