Replace the function inlining mechanism with new platform operators

[pavpanchekha]

In Herbie, you can write a named FPCore, as in (FPCore name (...) ...). Then, in later FPCores, you can write (name <expr>) to call that FPCore; you have to do the correct number of arguments and types and everything but it'll do what you expect. Currently, the way this works is a little hash table on the side called *functions* storing named FPCores and their definitions; when the desugarer sees one of these named functions, it'll inline the definition.

This PR switches functions to just use the platform feature. When Herbie sees a named FPCore, it'll add that an operator to the current platform, using the name for the :fpcore format and the spec and implementation for the operator spec and impl. Herbie will now generate direct calls to this operator, which is nice, and the rewriter still has access to the spec so can inline if it wants.

WIP:

• I don't think this works with places; it probably needs a re-register pass somehow for places, or we need to switch to parsing in places.
• I'm a bit worried that we'll generate self-calls or reverse-calls. Maybe that's fine, because right now we do expect humans to manually update the FPCore?

[pavpanchekha]

An update on this PR. It now seems correct. I have reviewed all but the big four files: read.rkt, platform-language.rkt, load-platform.rkt, and platform-state.rkt. I think the contents of those are correct but I am not thrilled with the internal APIs.

[pavpanchekha]

I'm now pretty happy with results, but the nightly fails. Will investigate, but I think we're close to merging.