kernel-compile benchmark has unstable prerequisite

[parthy]

The kernel-compile benchmark initrd takes a linux attribute as an argument to infer the tarball using the linux.src value. However, this is subject to change when Linux version changes (even a minor bump) and thus the results are not comparable anymore. Such a change should be made explicitly.

One option would be to pass the tarball directly to the initrd creation. I believe this could be possible right now by "faking" a linux argument that has linux.src pointing to an explicitly chosen Linux tarball, but it feels a little hacky to do that. And even that seems not far enough because all other packages in the system can still change, potentially skewing the results. Maybe the entire pkgs to be used in the initrd should be passed as a separate input to all initrds?

I am open for suggestions on how to make this benchmark easy to use and also allowing for a reproducible operation even when the global nixpkgs are updated.

[parthy]

What's more, the continuous output during the benchmark includes the entire output path in the measurement (e.g., printing to a serial console in some cases). I think the benchmark should be performed silently to eliminate those effects.