gh-150942: Speed up re.findall and re.sub/subn result building

[eendebakpt]

Append result items to the output list with _PyList_AppendTakeRef instead of PyList_Append followed by Py_DECREF, removing a reference-count round-trip per appended item (and a per-append lock on the free-threaded build).

Microbenchmarks

Benchmark	main	this PR	speedup
re_sub	1.93 ms	1.73 ms	1.11×
re_subn	1.92 ms	1.73 ms	1.11×
re_findall	1.89 ms	1.80 ms	1.05×
re_findall_groups	4.01 ms	3.87 ms	1.04×
geomean			1.06×

Benchmark script

"""Microbenchmark re.findall / split / sub / subn (result-list building)."""
import re
import pyperf

WORDS = "the quick brown fox jumps over the lazy dog " * 2000
CSVISH = "field1,field2,field3,field4,field5,field6,field7,field8\n" * 3000

word_re = re.compile(r"\w+")
group_re = re.compile(r"(\w)(\w+)")
split_re = re.compile(r"[,\n]")
sub_re = re.compile(r"\w+")


def b_findall():        return word_re.findall(WORDS)
def b_findall_groups(): return group_re.findall(WORDS)   # list of tuples
def b_sub():            return sub_re.sub("X", WORDS)
def b_subn():           return sub_re.subn("X", WORDS)


if __name__ == "__main__":
    runner = pyperf.Runner()
    runner.bench_func("re_findall", b_findall)
    runner.bench_func("re_findall_groups", b_findall_groups)
    runner.bench_func("re_sub", b_sub)
    runner.bench_func("re_subn", b_subn)

• Issue: Improve performance by using reference stealing methods #150942

[sergey-miryanov]

Looks good to me