Fix type inference for tuples in iterable context

[ilevkivskyi]

It is a minor change to fix the example in python/typeshed#7904

[ilevkivskyi]

cc @JukkaL

[ilevkivskyi]

Well, this shows how fragile is type inference that has neither backtracking nor full-expression scope. Here pytest uses a type like Iterable[Union[Tuple[str, Dict[str, int], str], Sequence[object]]] so after my first commit, a tuple of three item tuples, where one of items has second item unlucky to be a dict, there is an error, although such three item tuple would perfectly match Sequence[object].

I tried a fix (which btw makes things more consistent to be fair). If this will still cause problems as well, we can also try simplifying unions before providing it as a context for tuple.

[ilevkivskyi]

Actually, I am almost sure simplifying the union is the right thing to do. What if context is a union of two identical tuple types? in this case we currently give up and don't provide a context at all.

[ilevkivskyi]

OTOH if original union was like Union[Tuple[int, int], object], we will destroy the tuple by simplifying. Maybe we can try both original and simplified union and choose the one that has exactly one tuple-like type.

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

[ilevkivskyi]

Hm, OK, there is no effect in mypy_primer so maybe let's just not touch it. If we will discover some more issues later, we can try what I mentioned above.