Fix soundness bug by using currying information from typing

[stedolan]

Previously, transl_curried_function in Translcore redetected function currying, which is difficult with locals as modes may make it invalid to merge two lambdas into a single n-ary function.

The mode logic here was wrong, leading to a soundness bug. Rather than fix it (which would continue the duplication of mode-checking between typing and transl), the fix here is to add the relevant information to Typedtree, so that Translcore follows the decisions made by typing instead of redetecting currying on its own.

This causes a slight change in closure allocation / caml_apply behaviour, as the currying decisions made by Typedtree do not always exactly match those made by Lambda. For instance, the following is now detected as a two-argument function as the typechecker (unlike Lambda) does not see through the unboxed record:

type 'a unbox = A of 'a [@@unboxed]
let f a b = A (fun c -> a+b+c)

[lpw25]

Mostly looks good. One place that looks suspicious.

[lpw25]

This now looks safe, but I think it's too strict and doesn't always allocate things locally when it could.

[lpw25]

LGTM