Isolate sampling code from Herbie

[pavpanchekha]

This PR is attempting to isolate the sampling code (at least compile-specs but ideally also make-search-func and ival-eval) from Herbie infrastructure, so that it can eventually be moved into Rival. Long term, I think the goal is to have the following three main methods:

• rival-compile, an analog of make-search-func
• rival-apply, an analog of ival-eval
• rival-analyze, an analog of search-step

Isolating all this code will require at least:

• Remove dependence on operator-info
• Remove dependence on timeline
• Remove dependence on ulps-between
• Remove use of compile-specs in accelerator.rkt

[pavpanchekha]

Update: I've added all the abstraction layers I want, and made all of the Herbie methods we know and love/hate go through them. Right now there's a bit of confusing mis-layering in how preconditions are handled, but if that is fixed I think we'll have a pretty clean API and I can move most of the key methods into correct-round.rkt and start isolating in order to move things to Rival.

[pavpanchekha]

The most recent nightly has way more invalid 0 points and way more exit points. Plus, localize is way slower. That has to be fixed before merging, we can't have performance regressions! But at least it runs to completion, which is great.