Enter a description of the change.

use SIMD for I420ToARGB odd widths in a temporary buffer instead of using C for remainder.  Currently the C code does not exactly match the SIMD code, so an odd width produces different pixels than an even width, causing a subtle artifact.  By using SIMD consistently, there is no difference in even and odd widths.  Also the SIMD performance is faster, so even with overhead of memcpy, performance improves.

BUG=447
TESTED=out\release\libyuv_unittest.exe --gtest_filter=*I420ToARGB*
R=harryjin@google.com

Review URL: https://webrtc-codereview.appspot.com/55579004.