Allow errors for unsound Self definitions on non-final classes

[davidhalter]

This comes after #2148 which wants to make this a mandatory error.

But for now it should be totally fine to allow type checkers to error here, since it can clearly be shown to be unsound, see also:

https://discuss.python.org/t/disallow-access-to-class-variables-with-self/105434

I'm happy to revisit the original pull request, but for now I hope this should suffice. I will rerun the automated tests if and once no changes are required.

[davidhalter]

@JelleZijlstra Would you mind reviewing this? IMO it's a very uncontroversial change.

The problem is (quoting from discuss):

from typing import Self, assert_type

class ParentB:
    a: list[Self]
    @classmethod
    def method4(cls) -> None:
        cls.a = [cls()]

class ChildB(ParentB):
    @classmethod
    def method3(cls) -> None:
        assert_type(cls, type[Self])
        assert_type(cls.a, list[Self]) # This can be a different type, but is not in conformance tests
        assert_type(cls.a[0], Self)  # Same
        print(cls.a)

ParentB.method4()
ChildB.method3()  # Prints [<__main__.ParentB object>]

There is currently no part of the spec that describes this.

[JelleZijlstra]

I'd prefer to keep this in a separate PR now that we allow fixing type checker versions, I can prepare a PR now to update it though.

[davidhalter]

Makes sense, I'll try to revert it.

[davidhalter]

Zuban is now at the same version as before and the results are much more understandable