Miscellaneous performance improvements
I plan to work here on performance.
I may introduce breaking changes when I encounter a discutable design decision.
I got nice performance improvements in general but also some regressions in unicode-data. Some benchmarks are fragile and I will need to test on another machine. I think we should also use sample texts as in unicode-transforms.
Anyway, this PR is getting too big, so I am going to split it in various ones, starting with #109.
FYI, I am working on rebasing & splitting this PR into smaller PR. I will not make the mistake again with a huge PR. Leaving this open while not completely ported.
So long, big unmergeable PR.