scale to 3/4 or 3/8 with odd width destinations efficiently.  previously if width was not multiple of what the simd loop would do (24), scaling would fall back on slower C code.  This change allows SIMD to be used for most of the scaling and C for the remainder, improving efficiency.
BUG=314
TESTED=set LIBYUV_WIDTH=1896 & ScaleDownBy3by4_*
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/48249004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1380 16f28f9a-4ce2-e073-06de-1de4eb20be90