Removing divisions from od_dir_find8()

Instead of dividing the squared partial sums by the number n of pixels in the
line, we multiply by 840/n, where 840=3*5*7*8. This not only avoids the
divisions, but it also makes the optimization exact as there is no more
rounding.

ntt-short1 resuts:
         MEDIUM (%) HIGH (%)
    PSNR -0.012070 -0.059644
 PSNRHVS -0.016845 -0.020871
    SSIM -0.026984 -0.031257
FASTSSIM -0.026078 0.414901

Change-Id: Ie553d5e3a545dee860a00879d724ecfc00f0a974