Fixes #5365 and does some preparatory work for #174.

ECMAScript requires that all capture groups inside a loop are reset to "unmatched" at the start of each repetition. To do this correctly, we have to know the first capture group that appears inside the loop. We actually do not have to know the last one to reset the capture groups correctly: If a loop is repeated, then all capture groups between the first capture group inside the loop and the last one in the regex appear inside this loop or are already unmatched. (Knowing the last capture group inside the loop would avoid some unnecessary work, but given that the number of capture groups is usually low in practice it's not clear that this additional work is worth it when traded against more NFA traversal at the time of matching. This would require some benchmarking.)

To find the first capture group, the matcher now traverses the NFA until it encounters a capture group or has visited all of the nodes inside a loop. For POSIX regexes, the traversal is skipped; instead the first capture group is set to the number of capture groups in the whole expression, which means that no capture groups will be reset when processing loops as required by POSIX semantics.

To avoid this NFA traversal for each repetition, the result of this traversal is now cached in the loop state by extending the type _Loop_vals_t. But this means the type _Loop_vals_t and _Matcher have to be renamed for binary compatibility. At this opportunity, I also make some minor cleanups.

I split this PR into several commits to make reviewing easier:

• Rename _Matcher to _Matcher2 and add a template parameter for an allocator to prepare for <regex>: Should regex_match() use the match_results allocator? #174 as suggested by STL. For now, I'm passing void as the allocator argument because it's obviously not an allocator, so we don't have to worry about ABI when allocator support is actually added.
• Fix the type of the _Ncap member variable: It should be unsigned int, not int.
• Remove the member function _Get_ncap(): It has become superfluous because the type of _Ncap has been corrected.
• Remove the matcher's duplicate member variable _First in favor of _Begin (to match _End).
• Remove the unnecessary local variable _Cur_iter in _Do_rep to minimally reduce stack consumption.
• Actual fix of <regex>: Implementation divergence for capture group behavior #5365:
  • Extend _Loop_vals_t and rename it to _Loop_vals_v2_t.
  • Add function _Find_first_inner_capture_group that recursively traverses the NFA to find the first capture group. The recursive traversal is protected against stack overflow like all other recursive calls in the matcher.
  • Split original _Do_rep into _Do_rep (doing the main work for any repetition) and _Do_rep_first (doing the preparation before the first repetition).
  • Update _Do_rep to reset capture groups.
  • Update _Do_rep0 to reset capture groups in one for loop. (They already are reset correctly in the other for loop.)
  • Implement _Do_rep_first to perform the NFA traversal if necessary and update the loop state accordingly.