Re-commit r34072 with some modifications:
- turned off by default (set $wgAdvancedSearchHighlighting to turn on)
- reverted r26269, \b doesn't interact very good with unicode data, so it broke highlighting of words that end/begin in nonascii chars completely
- small bugfixes in unicode handling, tested in more languages
- $wgSearchHighlightBoundaries need to be set to "" for CJK wikis
- benchmarking: on typical simplewiki data, the code is around 4-5 slower (according to noc.wikimedia.org the old code profiles to about 0.8%), but can be up to 20 times slower on featured-size articles
- update release notes (also for r33400)
- fix profiling errors in SpecialSearch