[red-knot] remove wrong typevar attribute implementations

[carljm]

Summary

In #14182 I added "precise" per-TypeVar type inference for attributes of a typevar (__bound__, __constraints__, __default__). This wasn't motivated so much by the real-world need for this precise inference, as it was by wanting to test the correctness of our internal representation of the typevar, via an mdtest.

The implementation I added was unfortunately just wrong. The problem is that our internal representation cares about the bound/constraints/default inferred as a type expression, while the runtime attributes of the TypeVar object will return it as a value expression.

In the simplest case, this means that for a typevar [T: A], while we might store the bound as the Instance type A, the value of the __bound__ attribute should be typed as type[A].

Beguiled by the simple case, I slapped on a .to_meta_type() and called it a day. But this is just wrong in more complex cases; the meta-type is not always the same thing as the "type form" type. For example, the meta-type of A | B is type[A] | type[B], but the type-form type in this case would be an instance of types.UnionType.

In a future world with PEP 747 (TypeForm), where we have a typevar with a bound of A | B it would be correct for us to infer its __bound__ attribute as TypeForm[A | B], offering a nicer solution to this problem.

In the meantime, I think we should just follow the lead of other type checkers in falling back to the general typeshed types for these attributes, and I should bite the bullet and write a Rust test instead of an mdtest when I want to test internal representations of types that don't (yet, since we haven't implemented generics) have a visible effect in the type system.

Test Plan

Re-wrote an mdtest as a Rust test.

[github-actions]

ruff-ecosystem results

Linter (stable)

✅ ecosystem check detected no linter changes.

Linter (preview)

✅ ecosystem check detected no linter changes.