As per the linked bug report, the distance-weighted compound
prediction has two separate round operations, first by 3
bits (inside the various convolve functions), then by 10 bits
(after the convolution functions).

We can improve on this by right shifting by 3 bits inside the
convolve functions - this is equivalent to doing a single round
by 13 bits at the end.

Note: In the encoder, when doing joint_motion_search(), we do
things a bit differently: So that we can try modifying the two
"sides" of the prediction independently, we predict each side as
if it were a single prediction (including rounding), then blend
these single predictions together.

This is already an approximation to the "real" prediction, even
in the non-jnt-comp case. So we leave that code path as-is.

BUG=aomedia:1289

Change-Id: I9ad1fbcb3e12db2b5fc3c82b407f0fd9e6b39750