It didn't like passing reference to something that could be treated as a value.

Variables is noncopyable but default copy constructor was apparently working in VS.

Each subsequent rule was reusing variables object, which wasn't good.
Parsing of sequences had to be changed not to end with semi,
and it wasn't actually passing on lexicons in the first place.

It was looking for minimum value and Position::Nowhere is
the smallest you can get value-wise (min_int). So if there was
an empty Match in a vector, first_token for the vector would
always end up returning Position::Nowhere, which was wrong
(empty submatches should've been ignored and Position::Nowhere
returned only if there were no non-empty alternatives).

longest() and oneof() operator takes boost::shared_ptr<std::vector<boost::shared_ptr<ConjConditions> > >

So it's clear it's not about MatchRule at all.

Sort of a comeback of an old idea. It's either this,
or adding a third parameter soon for lexicons.
I went with this as this makes it easier to add yet
another thing to toss around, if such need ever arises.

1) MatchOperator was essentially a wrapper for "match" section of "apply" operator
2) Other operators like Repeat actually used ConjConditions directly, without using MatchOperator
   - so ApplyOperator could too!
3) The confusion was about "match" referring both to part that is matching
   sentence against match conditions, and the results of this matching process.
   That led e.g. to "match_operator" rule in grammar, which was unlike
   "strset_operator" rule and the like. "match_operator" was about
   the "match" section in "apply" and now it is more straightforward.
   Next step should rename "match_fit" to be actual "match_operator",
   consistent with "XXX_operator" for other XXX types.

Changing submatching syntax from "->" to ":"  and M, MA to be shorthand for ($m:_M):1 and ($m:_M):2, respectively.
Also, "M" is optional if followed by ":", "M:1" is same as ":1"