Improve interfacing with ChatGPT

[ninehusky]

See #56 for more details. Right now, ChatGPT is doing poorly (good for us, in a way), but it's doing a bit too poorly -- almost too bad to be believable, honestly.

I think concretely, here are the things we should do:

• First, seed ChatGPT to make things non-deterministic/reproducible.
• Then, do some sort of "constrained encoding" style tricks to get it to stop producing as many rules outside the given grammar (syntax).
• We should also attempt to convince it to disregard (some of) what it knows intrinsically about operators, so it can reason about stuff like Halide division being implemented as a / 0 = 0, for example.