Add support for extensible type attributes

[craigfe]

This adds support for attaching arbitrary metadata to type representations, in the form of a new AST node:

  | Attributes : { attrs : 'a Attribute.Map.t; attr_type : 'a t } -> 'a t

where an 'a Attribute.Map.t can store arbitrary data (indexed by the type 'a).

The motivation is to allow interpreters to attach custom info to typereps that is not generally useful (avoiding hard-coding this information directly in the typerep AST). For example, we currently have a variant case Boxed : 'a t -> 'a t that is just a flag for the binary encoders (and must be manually ignored by all the other derivers). The attribute mechanism allows type_binary.ml to define its own Boxed attribute to use instead.

Another use-case for this feature is in allowing interpreters to effectively extend Custom by attaching hand-implemented functions to typereps. As a demonstration, the new type_random.ml deriver now supports custom value generators (a limitation that was raised in #58). This isn't yet exposed to external users (since users can't yet define their own interpreters either).

The implementation of this feature uses Higher, and (for once) I think it's strictly necessary to support the heterogeneous map semantics. I've defined a helper functor that prevents this from spreading further than it needs to.

[mattiasdrp]

I think I really need this PR to finish mine (#57) because the current state of Repr makes it hard to have some functions that are not directly implemented but waiting in a functor.

[icristescu]

Really nice!

[Ngoguey42]

I always love to read OCaml black magic! :)

[craigfe]

Thanks all for the reviews 🙂 Merging now.