10x slower performance over ICU

[Manishearth]

@zbraniecki did some bemchmarking of unicode-normalization against ICU: unicode-org/icu4x#93 (comment)

We're 10 times slower, which doesn't seem great. Worth figuring out why; I bet it's the lookup speed.

Some low hanging fruit: qc_nfd and qc_nfk are just matches, we should follow what unicode-segmentation does and have cached binary searches (ideally, ones with an ascii special case so every space character doesn't reset the cache).