This follow-on to #1384 removes now-redundant information from the type environment.

My concern before starting down this road is that the information will be needed in the translation to Lambda. It's not right now because everything is a value. But that's due to change, of course. Yet then the layout information is really a property of the translation algorithm, and needn't be persisted all the way through the type checker.(*) So there's a possibility that we'd have to create a little transl_env type or something to compute the layouts before translating. This is appealing, though, because it simplifies type-checking recursive types (no need to update data that you're not storing) and eliminates duplicate information. And, if we ever start being able to handle e.g. type ('a : any) foo = { blah : int; quux : 'a; quux2 : 'a }, then we won't be able to store the layouts in the type structure anyway, because they will depend on the arguments.

(*) I still agree with the plan to keep seeking any layouts we need in the type checker. So maybe it all takes the form of a function that computes the inner layout information from the layouts of arguments, all produced in the type checker (to get e.g. good error messages). But I think it will be easier to produce such a thing when we really need it, as we're implementing structures that contain non-values.