Add ability to filter out 'very small' changes

[rylev]

Fixes #1006.

This introduces a new filter (on by default) which filters out any changes that are deemed to have a "very small" magnitude. You can find the definition for magnitude here.

[Mark-Simulacrum]

I think we should make sure that the comparison view always show any regressions (or improvemnts) we report in the summary comments / perf triage reports by default. Otherwise I think people will be confused. AFAICT at a quick skim, those comments currently do show 'very small' changes, so this would be a step away from that.

[Mark-Simulacrum]

Not sure how we should fix that, though -- I lean towards not showing very small changes in perf summaries, under the assumption that they are likely to be irrelevant in practice (same reason we're hiding them in comparison pages).

[rylev]

So we only show the largest individual changes for either improvements or regressions in the triage and PR summaries. Because of this, those changes are only going to be "very small" if there are a decent number of very small changes. So it is true that it's possible that the comparison page doesn't show the change but the summary does, but only in specific situations.

There's only one way to fully fix this IMO: Don't use very small changes at all in determining perf categorizations. Thoughts on this?

[Mark-Simulacrum]

Yeah, I think that's the way to go.

[rylev]

@Mark-Simulacrum I've gone ahead and filtered out very small changes from the categorization determination.

[Mark-Simulacrum]

Looks like CI is failing but otherwise looks good to me.

Let's merge this in and flip the default for new significance in a few days (I'll be going through the untriaged regression PRs and get a sense for how things look).