TST: making lowest-deps envs reproducible

[neutrinoceros]

I'm seeing errors on reproject + Python 3.11 that I'm willing to bet are effectively noise, assuming

1. this job is really meant to check that astropy-stable + reproject-stable still passes tests on a controlled env with old versions of their combined dependency tree.
2. the error is just showing with recent enough versions of numpy

If my understanding is correct, then the solution would be to make this, and similar jobs run on reproducible environments, where everything is resolved to the lowest (oldest) compatible version, with the exception of reproject and astropy themselves (still resolved to latest versions). This is up my alley, as I already did this for astropy's own CI already, but does require a bit of coordination.
In particular, I need confirmation that I'm interpreting the scope correctly. I will also give #44 priority over any attempts to modify how tests are run here.

xref astropy/astropy#18782

[astrofrog]

I will try and push #44 forward in the next day or so so we can see how things are working there.

[pllim]

Is this still relevant now that #44 is merged?

[neutrinoceros]

It is ! None of the strategy introduced in #44 produce intrinsicly reproducible environments, as they all favor "latest" versions, where the only meaningful differences are in wether they consider pre-releases and/or dev versions during resolution.
Within this framework, what I'm describing would take the form of a new variant, closer to what oldestdeps does in astropy's CI, though not necessarily named oldestdeps itself. It'd call it "lowest" after the resolution strategy of the same name.