stubtest: distinguish metaclass attributes from class attributes

[tungol]

If the runtime attribute of a class is actually from the metaclass, consider it to be MISSING at runtime.

This only occurs a couple times in the stdlib: it shows up when a descriptor is present on the metaclass but not the class, and we want to lie in the stub and say it's a thing on the class anyway.

I found this after noticing that enum.auto.__or__ had a comment that said it
didn't exist at runtime, but stubtest thought that it actually did. The issue is that on 3.10+, type.__or__ is defined for the purpose of Union types, and stubtest doesn't know the difference between type.__or__ and __or__ on the actual class.

Currently this matches on these things in typeshed's stdlib:

abc.ABCMeta.__abstractmethods__
builtins.object.__annotations__
enum.auto.__or__
enum.auto.__ror__
types.NotImplementedType.__call__

This MR doesn't resolve any allowlist entries for typeshed, and it doesn't create any new ones either, but should generate more accurate error messages in this particular edge case.

[JelleZijlstra]

Should != be is not?

And can you add a test that changes behavior with this patch?

[tungol]

I think you're right. I'll have a test ready in a little bit.

[tungol]

Here's a test. In order to get it to where that test would not generate the error without this change, I had to make a different change so that method-wrapper types can also be considered a function by is_probably_a_function. That's the type that type.__call__ is, and if there's a downside to making is_probably_a_function aware of those it wasn't immediately apparent to me.