Require perf-regression's to be -triaged to pass bors?

[Mark-Simulacrum]

Following up on rust-lang/rust#88832, it seems like a reasonable step may be to restrict bors merges to only happen after any existing perf regressions have been justified.

This might have the negative side-effect of partially discouraging perf runs, but that is hopefully unlikely. Our lack of strong follow up with PR authors + reviewers on perf justification post merge does make it harder to push for it to happen; it may be that some broader T-compiler policy work is necessary here.

[rylev]

On the social side of things, I think we should push in #t-compiler meetings for reviewers to run perf runs more often. I think the following mantra applies: "when in doubt whether this could cause a perf regression, do a perf run"

[Mark-Simulacrum]

I think we've done good work there. We just need to push on the follow up to perf runs. Part of that is likely providing better tooling for diagnosing regressions -- currently it's a bit of an opaque step if the perf.r-l.o UI doesn't show you anything.