gh-150819: Cache ConfigParser converter names per class

[gaborbernat]

Constructing a ConfigParser scans every attribute on the parser with dir() and matches each name against a regular expression to discover its get<type> converter methods. That discovery produces the same answer for every instance of a given parser class, yet it runs in full on each construction. Tools that read several config files, or build a fresh parser per file, pay this cost repeatedly — build backends, linters and test runners commonly create many parsers over a run.

This caches the discovered converter names per parser class and reuses them on later constructions, keyed on the exact type so a subclass that adds converters is handled correctly. The discovered set is identical to what the scan produces today.

Constructing a parser improves from 390 µs to 78.9 µs, 394% faster; constructing and reading a real setup.cfg end to end improves by 19%.

Benchmark	base	patched
ConfigParser() construct	390 µs	78.9 µs: 394% faster
ConfigParser + read real setup.cfg	2.19 ms	1.84 ms: 19% faster

Benchmark (pyperf)

Run base vs patched by swapping Lib/configparser.py on the same interpreter. The config text is a real setup.cfg from the top-1000 corpus (msal).

import configparser, pyperf

cfg = """\
[metadata]
name = msal
version = attr: msal.__version__
license = MIT
classifiers =
    Development Status :: 5 - Production/Stable
    Programming Language :: Python :: 3
[options]
python_requires = >=3.9
install_requires =
    requests>=2.0.0,<3
    PyJWT[crypto]>=1.0.0,<3
    cryptography>=2.5,<51
[options.extras_require]
broker =
    pymsalruntime>=0.20,<0.21; platform_system=='Windows'
"""

runner = pyperf.Runner()
runner.bench_func("ConfigParser() construct",
                  lambda: [configparser.ConfigParser() for _ in range(20)])

def parse_real():
    for _ in range(20):
        cp = configparser.ConfigParser()
        cp.read_string(cfg)
runner.bench_func("ConfigParser + read real setup.cfg", parse_real)

Resolves #150819.