split schemas

[aaazzam]

kitchen sink of previous PRs to benchmark

[codspeed-hq]

CodSpeed Performance Report

Merging #14644 will not alter performance

_{Comparing split-schemas (fcb2d98) with main (afdb04d)}

Summary

✅ 5 untouched benchmarks

[aaazzam]

Okay, so tl;dr is we shaved 100ms off of a 1.9s import speed in dev. The tests that are failing is b/c of a long tail of weird object access patterns that should be cleaned up independent of this PR.

I think this is a game of inches, so I'm eager to take a W where we can on this. That said, this moves us to a really gross single-class-per-file pattern that @chrisguidry and I both hate.