Improve the test infrastructure

For the first two years this project has run off the C3D2 Hydra: https://hydra.hq.c3d2.de/jobset/c3d2/microvm.nix#tabs-jobs That has the following problems:

• Our aarch64 builders are underpowered/broken.
• We build and check only the main branch. It would be nice to have early feedback on each PR.
• checks and hydraJobs in flake.nix containing multiple NixOS configurations requires a lot of RAM for a nix flake show, too much for a Github workflow. That breaks publishing to flakehub.com and flakestry.dev.