perf(intl): avoid parser offset scans

[xHeaven]

The parser used mb_substr() every time it read the next character. With UTF-8 strings, mb_substr() has to scan from the start of the string to reach that character, so long translations got much slower than they should.
In my benchmark, these numbers showed up:

Characters	Original	New	Speedup
2,000	4.568 ms	0.275 ms	16.6×
4,000	17.252 ms	0.545 ms	31.7×
8,000	62.975 ms	1.082 ms	58.2×
16,000	239.825 ms	2.141 ms	112.0×
32,000	936.264 ms	4.258 ms	219.9×