bug fix for odd width 16/24 bit to i420

bug fix for odd width 16/24 bit to i420

A bug was introduced on arm when the code for 'any' width switch to
a temporary stack buffer and simd.
The C version handles odd width by doing 1 pixel, instead of averaging 2.
But the SIMD any version is supposed to replicate the last pixel, then
the subsampling in Neon will average the pixel with itself, producing
the same result.
The previous version did this, but only for ARGB 32 bit, which was to
avoid introducing issues with subsampled YUY2 source.  This CL adds
replication for RGB 16 bit values.

TBR=harryjin@google.com
BUG=libyuv:510
Committed: https://chromium.googlesource.com/libyuv/libyuv/+/90335f604395321eadae07157aede2a54da821de

[fbarchard1, 2015-10-22 01:23:05]

Committed patchset #1 (id:1) manually as
90335f604395321eadae07157aede2a54da821de (presubmit successful).