BENCH: consistently test benchmarks (specifically argmax/argmin)

[mattip]

In PR #17989 changes to unrelated code changed argmax/argmin benchmarks significantly. We should be more consistent about benchmarking PRs so we have a baseline. Perhaps the benchmark itself is unstable, perhaps somehow this PR changed some compiler branching. It would be nice to set up daily/weekly/per PR/ benchmark runs and compare them before merging PRs.

For this particular problem, we should try different dtypes and benchmark ordering to see what is causing the change in argmax/argmin.

[mattip]

I thought this might be a duplicate but I could not find an issue to more tightly integrate asv (which seems to have last run in 2020) or codespeed (used by cpython and pypy). Related issues: #15987 and #15992

[seberg]

After the new discussion on the mailing list, we should likely set some mallopt's when running benchmarks, just to remove that from the possible reasons for fluctuations.

Having the benchmarks run regularly/automatic again would be nice indeed. And if those random fluctuations go away even "regression notifications" could be nice to have again.
(I have no big interest in pushing this, but I would be open to ideas such as asking for a small-development grant to get a good system up and running.)

[Illviljan]

Inspired by https://labs.quansight.org/blog/2021/08/github-actions-benchmarks/ xarray has added a workflow that you probably can copy/paste pretty easily, see pydata/xarray#5796.