Get more consistent distributions in parallel scenarios test

[abadams]

Note that this is not denoising the individual runtimes. The point of the test is to measure a distribution of runtimes for a variety of scenarios. This PR takes more samples for each distribution, and spreads the samples out over time to dodge transitory OS stuff by interleaving different test scenarios.

I also tweaked some of the scenarios to be more useful, and reported actual deciles so that the median is actually there.

[steven-johnson]

ready to land?

[abadams]

Looks like there was a merge conflict to fix. Apart from that, assuming everything goes green, yes

[abadams]

Failures are all vulkan stuff