Conflict between structural subtyping and self annotated with a type variable

[nilehmann]

If I create a protocol with a method that uses a type variable to annotate both self and the first argument of the method, mypy will accept as subtype, classes with methods accepting as first argument a parameter of the same class type. This unrestricted behavior can be used to get around the type system.

T = TypeVar('T')
class Fooable(Protocol):
    def foo(self: T, a: T) -> int:
        ...
class A:
    a = 1
    def foo(self, a: 'A') -> int:
        return self.a + a.a
class B:
    b = 1
    def foo(self, b: 'B') -> int:
        return self.b + b.b
def foo(a: Fooable, b: Fooable) -> None:
    a.foo(b)
foo(A(), B())

[ilevkivskyi]

This is not something specific to protocols. Exactly the same problem exists for nominal classes that use self-types. So this is almost a duplicate of #2353. I think it not exactly a duplicate because the solution for protocols may be different (but likely we will just go with the option 4 proposed by @JukkaL).