`<regex>`: Avoid heap allocations for few capturing groups and stack frames

[muellerj2]

Even if string matching only involves a few capturing groups or repetitions, a few buffers are still allocated on the heap by the matcher:

• The vector<bool> _Grp_valid in _Matcher3::_Tgt_state to store which capturing groups are matched.
• The vector<_Grp_t> _Grps in _Matcher3::_Tgt_state to store the extents of capturing groups.
• If matching follows the leftmost-longest rule, vectors of the same sizes in _Matcher3::_Res.
• The stack frames in vector<_Rx_state_frame_t<_It>> _Matcher3::_Frames.

Only one of these vectors, _Matcher3::_Frames, can grow beyond its initial size during matching.

These heap allocations are comparatively costly especially when matching short inputs, so we should avoid performing these allocations and instead use some stack-allocated buffer(s) if the required amount of memory is small enough.

[muellerj2]

I think a natural solution is to allocate a single stack-based buffer, use this buffer to supply the memory for the capturing group vectors (if the memory required by each of them is small enough) and assign the remaining memory to the stack frame vector.