Provide better analysis assistance in the comparison page

[rylev]

We currently only offer minimal assistance in analysis on the comparison page:

• Noisy benchmarks are marked as such (though confusingly, it doesn't matter if one benchmark/incr-cache-state is noisy, the entire benchmark gets marked as noisy)
• Changes above +- 0.2% get marked in light green or red, and changes changes above +- 1.0% get marked in darker green or red.

Now that we have an automated understanding of benchmark noisiness, we can do a lot better. Here are some ideas:

• Flatten the benchmark data so that all benchmark/incr-cache-state pairs are shown in the same table without expanding
• By default, filter all "non-significant" changes. We need to decide what it means to be significant. We have various definitions.
• Make the green/red highlighting depend also on if the benchmark is noisy or not.
• Better expose how a particular benchmark is noisy (i.e., is the benchmark actually noisy or just highly variable)
• Show statistics of how many benchmarks showed significant changes in each direction.
• Show correlation between cache-state/profile and significant changes (i.e., if all the significant changes were in optimized builds, we should show this)