feature benchmark: repeated scenario runs

[nrainer-materialize]

This implements MaterializeInc/database-issues#8565.

Nightly

https://buildkite.com/materialize/nightly/builds?branch=nrainer-materialize%3Afeature-benchmark%2Frepeated-scenario-runs

[nrainer-materialize]

I still need to validate and test the changes.

[nrainer-materialize]

This causes way more flakes. We need to discuss at the onsite whether we want to proceed with this change and possibly increase thresholds or discard it.

[def-]

In my opinion we should keep feature-benchmark as is so we can keep catching regressions, and improve parallel-benchmark to be the one-run-reliable benchmarking framework. First issue about where parallel-benchmark was not entirely consistent: https://github.com/MaterializeInc/database-issues/issues/8571

[def-]

I'll rebase this on top of #29664, maybe that helps? https://buildkite.com/materialize/nightly/builds/9644
Edit: Didn't help

[nrainer-materialize]

In my opinion we should keep feature-benchmark as is so we can keep catching regressions

I changed this PR to still always conduct three runs per scenario but pick the best outcome (instead of the median outcome) with 7478e73.

[def-]

Good compromise