Handle interactions between recursive aliases and recursive instances

[ilevkivskyi]

This is a follow-up for #13297

The fix for infinite recursion is kind of simple, but it is hard to make inference infer something useful. Currently we handle all most common cases, but it is quite fragile (I however have few tricks left if people will complain about inference).

Btw a curious thing: it looks like "hm..." is a valid value for Nested = Sequence[Union[int, Nested]].

cc @JukkaL

[hauntsaninja]

Looks great. Yeah, that is curious, but it makes sense (pyright also does the same thing)

[ilevkivskyi]

I will add couple more cases with a bit more principled inference recursion detection. Btw there is a thing that complicates things that is not specific to recursive types (see e.g. #11149), when mypy sees things like Union[T, List[T]] :> List[int], it can't decide which one to choose T :> int or T :> List[int]. This can be solved by adding an upper bound for T which forces mypy to choose one. I expect this to appear more often for recursive types (as they are in 95% of cases union-heavy), so we may add workaround with upper bound to the "Common issues" section in the docs.

EDIT: fixed the direction when this causes problems

[ilevkivskyi]

I actually added a bit of special-casing to handle the above problem specifically for recursive types. It should handle most of simple cases. I don't want to add this in general, as constraints inference is a very sensitive place.

[github-actions]

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