[red-knot] Trust all symbols in stub files

[sharkdp]

Summary

Follow-up on the discussion here. Generally trust undeclared symbols in stubs, not just at the module level.

Test Plan

New Markdown test.

[github-actions]

mypy_primer results

No ecosystem changes detected ✅

[AlexWaygood]

Oh, I suppose a lot of class-level assignments in typeshed are in the context of enum classes (e.g. https://github.com/python/typeshed/blob/ae57d69c9354147bcf7086d6db47af8f8c82bf7b/stdlib/re.pyi#L201-L221 ), and we infer all attributes on enums as Todo anyway...

[sharkdp]

I found one example that we do infer differently on this branch! I'm sure it's widely used:

from pdb import Pdb

reveal_type(Pdb.do_b)  # Previously: Unknown | def do_break(self, arg: str, temporary: bool = EllipsisType) -> bool | None
                       # Now:        def do_break(…) -> …

[AlexWaygood]

This does still feel more consistent to me, so I do support it still!

[carljm]

I think this makes sense as well. Sorry I didn't get through my notifications fast enough to comment on the previous PR sooner!

The core "gradual guarantee" idea that we should not emit false positives on working untyped code, when the user hasn't declared their intentions through type annotations, just doesn't apply to stub files in the same way, since they are inherently somebody explicitly providing type information.