Refactor benchmark suite to collect statistically significant results

[lucaswoj]

Note that map-load benchmarks pretty much always has the branch time much lower than the master one, making this particular benchmark not very useful. Maybe it's because V8 warms up on the first run.

#3224 (comment)

cc @mourner @jfirebaugh

[jfirebaugh]

Can we use Benchmark.js? It handles repeating the benchmarked code until it's able to collect a statistically significant result.

[lucaswoj]

@jfirebaugh Sounds like a good fit! I'll look into it.

[lucaswoj]

After toying with this a little, I think using benchmark.js specifically is more trouble than its worth. benchmark.js does not work well with asynchronous setup / teardown or non-time-based tests. We should be able to implement our own stat sig metrics.