Be more careful about the sigma->radius conversion, blur propagation,
and don't forget to undo the pre-existing scale for the RenderEffect
impl. It seems that all implementations of blur we have now - mask
filter, CPU blur, and RenderEffect blur - look mostly the same.

Based on the old box blur algorithm with a running sum, write a new one
that is more readable and more performant, due mostly to avoiding the
use of Bitmap.getPixels() and Bitmap.setPixels().

The blur is still, well, boxy, which we will probably be able to fix by
stacking two invocations of it, or trying some trickery to approximate
the Guassian kernel better but still keeping the linear time complexity.

Seems to provide slightly better perf.