More incremental optimization on cleanUp():

More incremental optimization on cleanUp():
* when splitting ascii vs non-ascii chunks, don't split punctuation and control chars as aggressively; this benefits the Korean test data
* use output buffer and echo; it's _slightly_ faster than string concatenation.
* Separate the surrogate check from the others; many Korean letters fall in the adjacent area with the same head byte, so this gives a small speed boost on Korean text