Cmm.value_kind cleanup

[Ekdohibs]

Currently, Cmm.value_kind is used only for unboxing code, when creating the Cmm. Besides, it adds a lot of information which is redundant (with the machtype) and sometimes simply incorrect (when it puts Vval Pgenval as a default value).
This is a first possibility for removing it without sacrificing optimisations, by replacing it with a simpler type that either says we have a value with Lambda.value_kind, or that it can be anything.
However, I still do not like this idea of having these value_kind nodes everywhere in the AST while they are used only for building the Cmm and never after. Another option would be to change the code building the Cmm to keep the information about unboxing, but it might be a far larger change.
Perhaps the simplest option would be to revert #295, but instead add kind annotations on Flambda lets (which are increasingly needed due to the layouts changes), and get unboxing to work as with closure.

[Gbury]

This looks good. Just need some more comments (either in the .ml and/or the .mli) to clear any doubt a ready might have.

[Ekdohibs]

Note: there are very slight semantics changes, such as in this case:

type 'a t = A : float t | B : int t
let  f x =
  let[@local] aux (type a) (x : a t) : a =
    match x with
    | A -> 0.
    | B -> if opaque false then raise Exit else failwith ""
  in
  let y = aux x in
  t x +. 1.

Now, y is unboxed where it was not previously, because in the B case, we detect it is No_result, whereas it immediately returned No_unboxing before when encountering the if having a return type of int.