Fix performance regression on numerical libraries

[ilevkivskyi]

My recent PR #22007, that improved overload vs callable subtyping, introduced ~10% performance regression on pandas/numpy/etc. There is no point in simplifying unions in code that is not visible to the user. Also m^2 / 2 + n^2 / 2 >= mn for all m and n, so IIUC this should be strictly better now.

cc @JukkaL

[ilevkivskyi]

OK, the crash is good, it uncovers a flaw in callable unioning logic: We should not combine callables with unpacked kwargs, since although they look like they have same shape, they actually don't. I will hopefully fix later today or tomorrow.

[JukkaL]

This looks fine, but the original PR seems to have also regressed this (false negative):

from typing import Protocol, overload

class P(Protocol):
    def __call__(self, x: int | str) -> int | str: ...

@overload
def f(x: int) -> int: ...
@overload
def f(y: str) -> str: ...

p: P = f      # accepted, but shouldn't be
p(x="s")      # fails at runtime

[ilevkivskyi]

This looks fine, but the original PR seems to have also regressed this (false negative)

This was semi-intentional. This kind of subtyping usually appears in the context of method overrides (not in context of callback passing) were we currently use ignore_pos_arg_names=True, so rejecting it completely would be inconsistent with callable vs callable method overrides. That said however, it looks like it should be not hard to simply pass the ignore_pos_arg_names value here, I will do this.

[ilevkivskyi]

It looks like GitLab has some kind of outage

Cloning into 'cki-lib'...
fatal: unable to access 'https://gitlab.com/cki-project/cki-lib/': The requested URL returned error: 502

and consequently mypy_primer is broken. @hauntsaninja just FYI.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

scikit-learn (https://github.com/scikit-learn/scikit-learn)
- sklearn/externals/array_api_extra/testing/_testing.py:389: error: Unused "type: ignore" comment  [unused-ignore]

[ilevkivskyi]

The primer is good, it was a false negative (different positional argument name were ignored).