This now reuses `fn discriminant_ty` in project, removing
some code duplication. Doing so made me realize that
we previously had a disagreement about the discriminant
type of generators, with MIR using `u32` and codegen and
trait selection using `i32`.
We now always use `u32`.