Determine applicability of new interfaces for probabilistic programming

@DhairyaLGandhi @Keno

Hi all. I'm trying to determine the applicability of some of this work to some of my interests in PP.

I recently started exploring a technique which utilizes Mjolnir and a simple type system to perform an optimization at `@generated` expansion time. The optimization is dependent on the "semantics" of the inference operation which you are performing (which is currently expressed through a combination of type inference and a reaching analysis). [I've written a small abstract here.](https://femtomc.github.io/mrb_dynamic_specialization.pdf)

My usage of Mjolnir is sort of funky - and I'm guessing the original use case is not quite like this. It almost seems like the new compiler work applies - but I really don't understand it yet, so I can't say with certainty.

Nonetheless, I wanted to start this issue as a reference so that this topic might be explored (I know AD is being explored as we speak).

Start by reading the linked abstract on Mjolnir and dynamic specialization, then inspect the repository's new compiler interfaces and existing Mjolnir usage. Compare those interfaces with the described @generated expansion-time optimization and inference semantics; done means documenting whether the approach applies and what questions remain, including the relationship to ongoing AD work.

Written by the indexing model from the issue text.