Benchmarking suggests that replacing some of the conversions from float to int

Benchmarking suggests that replacing some of the conversions from float to int
with direct floating-point comparisons in computeTypeMask() is a significant win
(up to 40%?).
In practice, we see a 10-15% speedup in this function (the current most
expensive Skia function) in Chrome on the FishIETank demo.

The remainder of the function looks like it ought to be amenable to similar
speedups, but we haven't found a rewrite that behaves well in benchmarking yet.

http://codereview.appspot.com/4551097/

git-svn-id: http://skia.googlecode.com/svn/trunk@1512 2bbb7eff-a529-9590-31e7-b0007b416f81