Use our own splitting in normalize

`String.prototype.split` is slow with regexps. Writing out the C-style string
searching code makes us take .8x the time on the "iterate.already.parsed"
benchmark from the "wasm" branch.

With the old String.prototype.split:

    Samples     Total (ms)  Mean (ms)   Standard Deviation (ms)
    50          325864.18   6517.28     162.15

With the new custom splitting:

    Samples     Total (ms)  Mean (ms)   Standard Deviation (ms)
    50          257604.64   5152.09     221.19

Author: fitzgen