Given a selector list, ".a, .b", our selector parser would, walk the tree and
collect matches for .a, then walk the tree and collect matches for .b. For this
to work, it had to keep a dedupe list AND, in the end, the order wouldn't match
Chrome.
The iteration is now reversed and aligned with Chrome. We walk the tree once
and each selector against the current element.
Almost all cases are faster now, even single selectors (which does the same
amount of walking) because the hashmap has been replaced with an ArrayList.
The only slower case is a selector list where one of the items is an #id, since
it no longer goes through the optimized path (we could preserve part of the
optimization, but it would be difficult, it's uncommon, and the lost is very
small).
Also, the selector cache now uses the new-ish ClockCache.