gh-93910: micro optimise enum attribute access

[hauntsaninja]

This halves access to attributes such as methods or _sunder_ name

Benchmarked using:

from enum import Enum

class Color(Enum):
    RED = "Red"
    BLUE = "Blue"
    GREEN = "Green"

def f():
    for _ in range(1000):
        Color.RED
        Color.BLUE
        Color.GREEN

def g():
    for _ in range(1000):
        Color.RED._name_
        Color.BLUE._name_
        Color.GREEN._name_

import timeit
print(timeit.timeit('f()', number=10000, globals=globals()))
print(timeit.timeit('g()', number=10000, globals=globals()))

As of today morning:

6da988a46c λ ./python.exe bm_enum.py
1.499613167019561
1.5516562079428695

As of the latest enum change #94913:

c20186c397 λ ./python.exe bm_enum.py
1.4998537090141326
16.095267291995697

As of this PR:

gh-93910-micro λ ./python.exe bm_enum.py
1.4983433339511976
8.81327024998609

[hauntsaninja]

Hmm, actually given that _member_names_ is a list, this is a little expensive for moderate to large enums (equals current main at >50 members). So maybe we throw this change away.

Note that there's a micro-optimisation we can do where we preserve condition order, but don't get self_dict if we don't need it:

- if isinstance(value, cls) and name not in self_dict and name in self._member_names_:
+ if isinstance(value, cls) and name not in super().__getattribute__('__dict__') and name in self._member_names_:

This reduces benchmark time to 12s (from 16s on main, compared to 1.5s yesterday) and should be a strict improvement regardless of enum size.

[ethanfurman]

Thank you for the effort, but it has been decided to not use __getattribute__.