Get rid of the "flipping behaviour" in narrowign by equality/identity

[ilevkivskyi]

We have this code that makes sure that if the intersection of declared and narrowed is not representable, then we prefer the narrowed. I added this logic a while ago, since this is what a user usually expects in case of e.g. an assignment or even an isinstance() check.

However, the recent effort to make equality/identity narrowing symmetric also caused this to result in a "flipping" behavior, when sides exchange types after a comparison, see e.g. testNarrowGenericCallableEquality. Although such cases are rare, I think the resulting behavior is really weird/confusing.

So I propose to add a flag to each restriction in binder (similar to from_assignment), say prefer_declared that would be set to True for right-most operand in equality/identity. This flag would override the fallback mentioned above. So that after something like assert x == y == a, the type of a will not change if the intersection is not representable.

I understand this breaks the symmetry, but this is purely practical thing, this is why people write x == 2 and not 2 == x.

cc @hauntsaninja

[ilevkivskyi]

Also to be clear, if people agree here, I volunteer to do the work.

[hauntsaninja]

Sounds worth trying to me! I don't have a strong philosophical stance here. I'm fine with the symmetry break, I also think the symmetric flipping doesn't seem bad in practice. Or I guess maybe my actual philosophical stance here is that in the long run we will have more intersection types :-)