Support A-normalization, CSE & cleanups

[Blaisorblade]

• Support A-normalization, written in CPS. In fact, Tillmann convinced me we probably want to use a writer monad instead of CPS, because it's more flexible for variations of A-normalization (such as producing one let with many variables instead of many lets, and relatedly CSE).
• Support A-normalization using mutable state & CSE
• Support let in untyped syntax & type inference.
• Cleanups in type inference.
• Tests.
• Lots more stuff which I noticed in the meantime.

Proper tests now in & API cleaned up.

[Blaisorblade]

Next goal: integration testing — applying this on the existing examples, for instance. But that should be in a separate branch. Hence, merging this.

[Blaisorblade]

This saves partial applications as intermediate results, and I've just realized this is not just strange but actively wasteful.

EDIT: in some cases, that does save work (when the partial application can in fact reduce), but if so wants the programmer, maybe he should manually share the result. I think the issue is complicated and for GHC there's some work on handling this.

[Blaisorblade]

This saves partial applications as intermediate results, and I've just realized this is not just strange but actively wasteful.

In fact, this part is easy to change, see aa839fde565453f8e6f06c88018058ec6290d4da and #2. The current behavior is "wasteful" if we also save all variables bound in A-normal form, as needed for static incrementalization. However, in that case it's less clear what one should do, as discussed in the past.