[MLton] ref flattening enabled

[Stephen Weeks]

> I don't know if it is a big deal or not about box being a constructor.  I
> don't think (but I certainly haven't tested it out) if I often apply
> constructors and then count on the result still being polymorphic, other
> than perhaps tuple or something.  Still, if it isn't hard, why not.

It's a little harder, but not too bad.  I'll just have the front end
expand the constructor away, much like we do for refs.