Mapping overloads cause incorrect variance inference

https://mypy-play.net/?gist=42b235155079dd44a79e66cbcffb0060

from collections.abc import Mapping

class A: ...
class B(A): ...

def test_mapping() -> None:
    class MyMapping[K, V](Mapping[K, V]): ...

    def _0[K](arg: Mapping[K, B]) -> Mapping[K, A]: return arg
    def _1[K](arg: MyMapping[K, B]) -> MyMapping[K, A]: return arg  # ❌️

mypy doesn't seem to reuse known variance of parent classes, but instead rechecks everything. For mapping, this breaks and mypy incorrectly infers invariance in the value type because of:

https://github.com/python/typeshed/blob/a9fbf47e676865bd5bcd94e121624d9e4b53c882/stdlib/typing.pyi#L787-L788

While I do think mypy could improve this behavior on their side, we really, really should avoid using covariant parameters in contravariant positions, especially with the knowledge that it can screw up variance inference when using 3.12 style annotations.

Start by reproducing the supplied example in the mypy playground, then inspect the Mapping definitions at the linked typeshed typing.pyi lines 787-788. Determine how their variance annotations interact with the 3.12-style MyMapping example; done means the example no longer produces the incorrect variance error without introducing an equivalent annotation issue.