Coq formalization of IDL-Soundness

[nomeata]

based on our experience, we really can really use some formal treatment of our
Candid work.

This branch contains some initial work. Current status and rough plan:

• A simple Coq setup (using the more modern dune-based setup)
• A nix-setup to build this
• Simple CI integration
• A Coq-ification of the definitions in IDL-Soundness.md
• Mechanizing the “canonical subtyping is sound” proof in that document
• Wait for the discussion in Still not sound: Gradual typing vs. opportunistic decoding #141 to settle
• Verify the rules of a “toy candid” (with much fewer types, but still
  enough to express the interesting bits)

[rossberg]

He he, thumbs up!

[nomeata]

This might be good enough to get into the repo. Not tacking Candid proper yet, that should wait for #141 to settle.