Hyrex runs on CPU!

[cgiovanetti]

Big refactor of the background calculation to fix #2. Now it happens in 3 steps:

1. compute H, T, nH, etc--smooth background needed for HyRex. Can occur on GPU if one is available.
2. compute recombination history with HyRex, pinned to CPU
3. integrate conformal time, visibility, etc. with pre-recombination background inheritance + HyRex output. New Background module has everything the old one did so we don't need to propagate downstream.

I see a ~2s speedup on the Perlmutter A100s.

In addition this fixes #25, by adding a new specs parameter that controls the relevant VeryChord tolerance. If the user requests reverseAD with checkpointing, this parameter is automatically adjusted.