CI benchmarking enhancements

[ian-r-rose]

A couple of small issues we have seen in the last few days which would improve the DX of running CI here:

• If the artifacts from a previous CI run persist artifacts, that can confuse the process-results step on rerunning, cf: https://github.com/coiled/coiled-runtime/runs/7867069146?check_suite_focus=true
• If there is a regression detected, we don't get to see the static site (you can download the results and generate it yourself, but that's a pain). Instead, we should generate the site, upload it as an artifact, and only upload to gh-pages on main.