Use combined three pass code for image blur.

This changes more closely matches the GL output, and the runtimes are similar or
faster for the common cases.

x86_64 times:
benchmark                            old-Us   new-Us   old/new
blur_image_filter_large_80.00_80.00  4842.04  2626.10  1.84381
blur_image_filter_small_80.00_80.00  3297.72  854.97   3.85712
blur_image_filter_large_10.00_10.00  930.44   720.50   1.29138
blur_image_filter_small_10.00_10.00  69.96    42.15    1.65979
blur_image_filter_large_1.00_1.00    682.66   521.78   1.30833
blur_image_filter_small_1.00_1.00    19.21    14.43    1.33125
blur_image_filter_large_0.50_0.50    696.17   64.14    10.8539
blur_image_filter_small_0.50_0.50    16.26    5.02     3.23904

arm64 times:
benchmark                            old-Us    new-Us    old/new
blur_image_filter_large_80.00_80.00  42144.53  14128.42  2.98296
blur_image_filter_small_80.00_80.00  24840.58  4392.58   5.65512
blur_image_filter_large_10.00_10.00  3556.40   3793.70   0.937449
blur_image_filter_small_10.00_10.00  282.53    220.62    1.28062
blur_image_filter_large_1.00_1.00    2502.20   2937.99   0.851671
blur_image_filter_small_1.00_1.00    83.32     81.93     1.01697
blur_image_filter_large_0.50_0.50    5643.80   272.83    20.6861
blur_image_filter_small_0.50_0.50    141.02    38.29     3.68295

Cq-Include-Trybots: skia.primary:Test-Debian9-Clang-GCE-CPU-AVX2-x86_64-Release-All-SKNX_NO_SIMD
Change-Id: Ic53b3186607d5485477b92e4ca7b092bf1366c52
Reviewed-on: https://skia-review.googlesource.com/52771
Commit-Queue: Herb Derby <herb@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>
2 files changed