Use the get_attribute hook for dynamic classes

[lincolnq]

Whenever we find __getattr__ or __getattribute__ on an instance, the getattribute hook is called to give the hook a chance to modify the return type of one of those functions.

Closes #6259, #5910

(Note: Currently there are no tests. I can add them if important.)

[ilevkivskyi]

Thanks for PR! I have couple comments. Could you also please add tests? (You can look at the existing tests for plugin hooks to see how to add them.)

[lincolnq]

@ilevkivskyi Thanks for the review! I made the changes you requested.

One of the test cases is that the attribute hook can produce an error. I invented my own error, "Field does not exist", but ideally I would test that you can fallback to the normal mypy error handling -- but I don't think you can access a MemberContext to call mx.msg.has_no_attr. Is the way I did it ok?

[ilevkivskyi]

Thanks for updates!

Sorry, there is one more thing, could you please also update the docstring for get_attribute_hook() in mypy/plugin.py?

[lincolnq]

Hm, ~~the old docstring was wrong too~~ (it specifies that the hook will be called with the base class name where the field is defined, but it is called with the derived class name). I assume the documented behavior is the desirable one? [edit: not sure if old docstring was wrong anymore, investigating.]

[ilevkivskyi]

I assume the documented behavior is the desirable one?

Yes, I think giving the name of the base class where __getattr__() was defined would be consistent. So you probably need to update your code. (The fix for the old test I posted above.)