perf: skip ascii chunks in wide

Large inline JSON and CSS with a few non-ASCII characters went through
the slow width loop byte by byte, about 900 KB per copy of the pages
corpus. Whole chunks of printable ASCII are now counted at once.