Refactor write_benchmark_json to comply with github-action-benchmark format and ensure time series creation

[thehrh]

Turns out the workflow failed because the script had created dictionary entries when only a simple list of results was allowed.

• I've applied the fix in my fork and here are the plots after a successful workflow: https://thehrh.github.io/pisa-1/dev/bench (benchmarks will show up in this repository under https://icecube.github.io/pisa/dev/bench)
• An example set of .json artifacts whose data are included in the plots is available from here

Another problem which appeared in my branch was the automatic removal of the previous benchmarking results in the gh-pages branch by the documentation workflow.

As a result, I've adapted the documentation workflow, on the one hand, to output to docs, not the root directory, of gh-pages, and, more importantly, configured it to keep_files. This way, dev/bench/data.js will not get deleted and instead will be extended.

It just means we will need to set the PISA (docs) URL in the right column to icecube.github.io/pisa/docs (leaving the Pages source at / (root) in settings.

Also, ideally before this PR is merged, we should manually delete the gh-pages branch once so that the files currently in its root directory are removed.

In the future, the benchmark time series will then be extended as long as dev/bench/data.js in gh-pages isn't deleted. If —for some reason—it does get deleted at some point without there being a backup, the time series will just start anew.

Finally, an entry point for the benchmarking script is added to setup.py.