Fix record accessors for higher-rank field types

[runarorama]

A record-style data declaration whose field has a rank-N type — e.g.

type Box = { unbox : forall a. a -> a }

failed to typecheck. The auto-generated getter cases Box f -> f has no type annotation, so synthesis infers Box -> a2 -> a1 (the inner forall instantiates to distinct existentials that can't be re-generalized in a purely inferred result), and the typechecker complains:

I found a value of type:  a2 -> a1
where I expected to find: a -> a

The setter and modifier hit the same issue from the other side (constructor application with a rank-N argument).

The fix is to give each auto-generated accessor an explicit type annotation built from the data declaration's tyvars and the field's type. Check-direction Skolemization handles the nested foralls correctly, so the same accessor body that previously failed inference now type-checks cleanly.

Each field's annotation is shaped as:

getter:  forall ddTyvars. T ddTyvars -> fieldType
setter:  forall ddTyvars. fieldType -> T ddTyvars -> T ddTyvars
modifier:
         forall ddTyvars. (fieldType -> fieldType)
                          -> T ddTyvars -> T ddTyvars

The field types come from the /resolved/ data declaration's constructor (which already has every type reference bound to the namespace) rather than the raw parsed types — otherwise an annotation built from a parsed field type referencing another type in the same file would surface as "I don't know about the type X" because that reference hasn't been resolved at accessor-generation time.

Fixes #498.

[pchiusano]

I could have sworn I added a review of this but maybe I left it in drafts on my phone. Anyway, while I haven't reviewed the code in detail, this could use a transcript test to exercise, and it should illustrate being able to change the type of a field that is the sole field referencing some type variable(s). I remember this being the tricky part of writing out the fully general type.