Roughly. instead of 4 loads and 8 multiples, use 1 load and 2 multiples
4 times over.  The original code, as with the C code from clang and gcc,
did all the loads, then all the math, then the store.  The new code
does a load, then the math, then the next load, etc.
This schedules better on current arm 64 cpus.
Number of registers also reduced, reusing the same registers.

HiSilicon ARM A73:

Now
TestGaussRow_Opt (890 ms)
TestGaussCol_Opt (571 ms)

Was
TestGaussRow_Opt (1061 ms)
TestGaussCol_Opt (595 ms)

Qualcomm 821 (Pixel):

Now
TestGaussRow_Opt (571 ms)
TestGaussCol_Opt (474 ms)

Was
TestGaussRow_Opt (751 ms)
TestGaussCol_Opt (520 ms)

TBR=kjellander@chromium.org
BUG=libyuv:719
TEST=LibYUVPlanarTest.TestGaussRow_Opt

Reviewed-on: https://chromium-review.googlesource.com/627478Reviewed-by: Cheng Wang <wangcheng@google.com>
Reviewed-by: Frank Barchard <fbarchard@google.com>
Change-Id: I5ec81191d460801f0d4a89f0384f89925ff036de
Reviewed-on: https://chromium-review.googlesource.com/634448
Commit-Queue: Frank Barchard <fbarchard@google.com>