Create some benchmarks for fluent.runtime

[Pike]

@spookylukey has some benchmarks on his compiler branch, I'd love to land some variant of that ahead of time. That way we have a baseline idea on how the current code performs.

I'm wondering if we should, perhaps at least in addition to what's on branch, have some less micro-benchmark-y mini-benchmark, with, say 20 strings, 1 Term reference, one plural and variable ref? Or something like that? Still pretty unrealistic at that, but maybe a tad better.

And, maybe adjust the readme to use py-spy instead of pyflame? py-spy runs on more OSes, and bundles the flamegraph stuff.

[Pike]

Looking into this. My WIP is at 10 strings, and I see a ton of things that are not the resolver.

One thing is that timezone support in babel takes a long time to import. Even if I have a 10k loop around my 10 string resolver template, the initial babel import shows up. Roughly, importing babel timezone data on my machine is same order of magnitude as 10k string lookups, on python 3.7.

It's hard to drill down further into the balances between caches, because babel data really dominates.

I wonder if we should load babel stuff on-demand? In particular the date formatting, and plural rules.