Expand SizeAnalysis into a full simplifier pass

[rrnewton]

I can't think of a good way to separate these two. This pass should do standard partial evaluation / delta-reduction, and should also label the known sizes that result.

Right now it only gets the easy stuff. It should be hooked up to the interpreter so that it can go further.

In fact, we should find that the combinator-based representation of arrays should be better for static evaluation. For example, lookups into arrays created by generate should often be statically computable if that is needed to compute the shape of something else. [Still, we have to avoid executing the whole Accelerate program, because it actually is a complete, ready-to-run program.]

The other thing to be careful of is that we need to preclude having an effective caching system to avoid recompilation.