Remove invalid unwrap annotation

[dillonkearns]

Say we are writing codegen for a type like:

type MyType = Wrapper String

Elm.unwrap and Elm.unwrapper generate a lambda that pattern matches on a single-variant custom type:

 \(Wrapper val) -> val

The argument pattern matches on Wrapper, so the argument type should be the Wrapper variant's type, not a generic type variable (Elm gives an error if you use a generic type variable):

Before:

extract : val -> unwrapped
extract (Wrapper val) =
    val

TYPE MISMATCH: The argument is `MyType`, but the type annotation
says `val`.

Since we don't know the type name (MyType in the example I gave above) we can't infer it so we short-circuit the type annotation by returning Err [].

After:

extract (Wrapper val) =
    val

If you know the types, pipe through Elm.withType to attach them:

Elm.unwrapper [] "Wrapper" |> Elm.withType (Type.function [Type.named [] "MyType"] Type.string)

Returning Err [] from unwrapper would have lost inference for outer expressions too, since applyType used to short-circuit on any Err. For example, Elm.Op.plus (Elm.int 1) (Elm.unwrap [] "Wrapper" wrapped) would produce no signature at all. So applyType now distinguishes Err [] (intentionally unknown) from Err [errors] (an actual inference failure). When the function annotation is Err [], it synthesizes a fresh generic return type so downstream inference still flows, and the outer declaration still gets foo : Int.

If you would prefer to do a bigger refactor to have a custom type to represent an unknown inference value instead of having special meaning for Err [], we could totally do that too, just let me know!