Refactor benchmarking process to run relative benchmark for PR

[pllim]

Perhaps we can be inpired by https://labs.quansight.org/blog/2021/08/github-actions-benchmarks/ .

Looks like it has been deployed in scikit-image:

• https://github.com/scikit-image/scikit-image/pulls?q=label%3Arun-benchmark+is%3Apr
• https://github.com/scikit-image/scikit-image/tree/main/benchmarks
  • https://github.com/scikit-image/scikit-image/blob/main/benchmarks/README_CI.md
• https://github.com/scikit-image/scikit-image/blob/main/.github/workflows/benchmarks.yml

Also see:

• Move benchmark codes to astropy core lib and only keep results here #102
• Improving Python benchmarking tooling airspeed-velocity/asv#1219

[pllim]

For the GitHub Actions route, I think https://github.com/scikit-image/scikit-image/blob/main/benchmarks/README_CI.md sums up the expectations. We will never get an absolute number, only will know if a change introduced in a PR has changed performance significantly or not. So, if you are looking for absolutes (e.g., exactly how long is needed to run something and how that changes over time) or comparison with delta smaller than sensitivity of the CI, then this is not for you.

[pllim]

Done in astropy/astropy#15779 thanks to @MridulS !