Improve the summary message both in GitHub and in the triage report

[rylev]

This improves the summaries generated for both the perf bot on GitHub as well as the weekly triage report.

Note: I've not tested this locally as I don't have a local database in a useable form anymore.

Perf Bot

before

Summary: This change led to large relevant regressions 😿 in compiler performance.

• Large regression in instruction counts (up to 2.2% on incr-patched: reverse builds of regex check)

after

Summary: This benchmark run shows 10 relevant regressions 😿 to instruction counts.

• Average relevant regression: 1.7%
• Largest relevant regression in instruction count: 2.2% on incr-patched: reverse builds of regex check

Triage Report

before

• Very large regression in instruction counts (up to 8.2% on full builds of keccak check)

after

• Largest relevant regression in instruction counts: 8.2% on full builds of keccak check

Other Changes

Additionally, this PR changes the message slightly when there are no "relevant" changes (and thus nothing needs to be done) but there are still some (small) statistically significant changes. The following is added in the perf bot message.

"N results were found to be statistically significant but too small to be relevant."

[Mark-Simulacrum]

Feel free to merge with nit fixed

[nnethercote]

Oh! I've been thinking about this stuff, and have been meaning to make some changes myself.

• I would like the message to say how many relevant improvements/unchangeds/regressions there were.
• I would like finer classification of "mixed results" cases. Sometimes they are genuinely mixed, e.g. half improvements and half regressions. But sometimes they are like 90% improvements and 10% regressions, or vice versa. Adding a "mostly positive" or "mostly negative" classification might help here? Though perhaps the previous point would reduce the need for this.

What do you think?

[rylev]

• I would like the message to say how many relevant improvements/unchangeds/regressions there were.

I believe this is what I already have (it shows the number of improvements and regressions) except it doesn't say the number of "unchanged". Do you want the number of unchanged to be the number of test cases where there was no significant change - this is essentially just all other test cases that aren't improvements or regressions.

• I would like finer classification of "mixed results" cases. Sometimes they are genuinely mixed, e.g. half improvements and half regressions. But sometimes they are like 90% improvements and 10% regressions, or vice versa. Adding a "mostly positive" or "mostly negative" classification might help here? Though perhaps the previous point would reduce the need for this.

We could do this, but it might be useful to first see how these changes look in practice before figuring out the line between "genuinely mixed" and "mixed but leaning one way predominantly". What do you think?

[camelid]

Maybe "important" would be a better word than "relevant"? The distinction between "significant" and "relevant" seems a bit fuzzy, but "significant" and "important" seem better separated.

[nnethercote]

Do you want the number of unchanged to be the number of test cases where there was no significant change - this is essentially just all other test cases that aren't improvements or regressions.

I want this in the summary, basically: