Remove branching and nested Refs

[samestep]

Over the past couple weeks @ravenrothkopf and I worked through automatic differentiation on a bunch of examples, and we realized a couple things:

• As the Dex paper notes, we need a Read kind of Ref to serve as the transpose for Accum.
• In contrast to Dex, we actually don't need a constraint stating that a type is a Vector space; instead, when we accumulate, we can just take an existing value of any type to choose the topology, and then work inside the dynamically-defined topology of that value. This will become more important when we Support sum types #65 but for now it just means that Bool/Fin/etc stay as-is instead of being transformed into Unit.
• We can't handle branching until we Support sum types #65, so instead for now we're just replacing If with a Select expression just like Wasm's.
• Dex doesn't handle nested Refs, and we can't either; to deal with this, we need to allow functions to take multiple parameters.

This PR makes the necessary changes to our codebase to accommodate those IR changes, and also simplifes the IR in a couple other ways:

• Every instance of Vec<T> is replaced with Box<[T]> to save space, because the former stores three pointers while the latter only uses two; the exception is Ty which we're leaving as Vec<T> for now because ts-rs does not support Box<[T]> or Rc<[T]>.
• No more Func and Block. Each Call now directly encodes the generic types at that callsite, and each For, Read, and Accum now directly owns its body.
• The Scope type now explicitly says what Constraint it satisfies, because otherwise typechecking the IR would be more annoying.

[ravenrothkopf]

LGTM! I need to go back in and change the pprint stuff to support select tho so don't review the PR yet!

[samestep]

Thanks for reviewing @ravenrothkopf! Let's hold off on merging this for now, though; from our discussion earlier today it seems that we may not want to support nested Refs at all, in which case we need to adjust the IR a bit more.

[ravenrothkopf]

nice nice nice! thank you so much for this