This makes the intention of inherent vs. trait impls somewhat more
clear and also fixes (?) an issue where trait impls with an unresolved
trait were added as inherent impls instead (hence the test changes).
If the first type parameter gets inferred, that's still not handled correctly;
it'll require some more refactoring: E.g. if we have `Thing<T, F=fn() -> T>` and
then instantiate `Thing<_>`, that gets turned into `Thing<_, fn() -> _>` before
the `_` is instantiated into a type variable -- so afterwards, we have two type
variables without any connection to each other.
This is working, but I'm not that happy with how the lowering works. We might
need an additional representation between `TypeRef` and `Ty` where names are
resolved and `impl Trait` bounds are separated out, but things like inference
variables don't exist and `impl Trait` is always represented the same
way.
Also note that this doesn't implement correct handling of RPIT *inside* the
function (which involves turning the `impl Trait`s into variables and creating
obligations for them). That intermediate representation might help there as
well.
It improves compile time in `--release` mode quite a bit, it doesn't
really slow things down and, conceptually, it seems closer to what we
want the physical architecture to look like (we don't want to
monomorphise EVERYTHING in a single leaf crate).
