This change adds `Arbitrary` instances to the key IR data types (both as
syntax and as native IR), as well as evaluator functions for both. This
way, we can ensure that the evaluation of one version of the IR is
observationally equivalent to another version of the IR, or even a later
IR. It also adds a similar ability through both static file compilation
and the JIT, to ensure that the translation through Cranelift and our
runtime works as expected.
This actually found a couple issues in its creation, and I hope is
helpful extensions into more interesting programs.