<regex>: Optimize matching of branchless loops

[muellerj2]

Towards #5957. This implements recognition of branchless but non-simple (=reentrant) loops and expands the stack growth optimization in the matcher to such loops. The main change is to recognize such loops in the parser and make it set a new flag on such loops; the changes to the matcher are minimal, just requiring two line changes: One to teach it about the flag and the other one to teach the code in the _N_end_rep handler for simple/branchless loops about a different opcode used in the _N_rep handler for reentrant greedy loops.

This PR does not implement the corresponding change for non-reentrant loops. It would be easy to do so, since the updated _Compute_loop_simplicity() already recognizes such loops. But I'm skeptical that the optimization for non-reentrant loops is actually worth it: It avoids a few assignments in exchange for a branch and more code. For this reason, I would like to do some benchmarking first before I allocate a flag bit to non-reentrancy.

Benchmark:

benchmark	before [ns]	after [ns]	speedup
bm_match_sequence_of_9a1b/"(?:a*b)*"/100	24553.6	15346	1.60
bm_match_sequence_of_9a1b/"(?:a*b)*"/200	73242.2	32470.6	2.26
bm_match_sequence_of_9a1b/"(?:a*b)*"/400	111607	73242.2	1.52