sf: optimize luma sampling code

Removing some math from the luma-sampling
code reduces its running time by about half
(~.5ms -> ~.25ms at 1.1MHz on a little core
in some experiments).

Bug: 127973145
Bug: 134922154
Test: trace and compare sampleArea times when
      scrolling in news app

Change-Id: Ie53d9595bea6685cf45f53972b42daa5e32fcc8e