Benchmark suite(s)

[masklinn]

Ideas like #31 or #1 may or may not pan out but it's difficult to know without something objective which can be run repeatedly.

#32 improved running the FilteredRE2/regex-filtered benchmark programs, and that is very useful, but these are:

• pretty coarse (though realistic which is very useful for e.g. profiling, maybe)
• not self-measuring (requiring BSD time's -l or GNU time's -v and those are not trivial to parse the output of which is frustrating)
• only inclusive of the regex part, not the extraction/conversion
• only on devices by default, even though as it turns out UAs and OSs can have different enough workloads that benching them separately can be worth doing (cf Re2 bench updates, revert #31 for UAs and OS #32)

So we'd need:

• the existing scripts to allow running over all categories
• maybe make atom length tunable
• a benchmark of extraction over the sample (possibly using a mapping in order to not pay for the regex querying / lookup?)
• finding and benchmarking specific cases via criterion or divan, for both regex filtering and extraction, e.g. cases which are very costly (and maybe cases which are very cheap too?)