Re-design the compound mode reference motion vector fetch. Use
a single run to provide all the compound ref mvs. Save the potential
additional two ref mv search runs on single reference frames.

Tested on night_720p 50 frames at 800 kbps. The average time cost
on find_mv_refs calls is reduced by 15% (average 69875 us ->
60473 us). The overall compression performance change is less than
0.01%.

BUG=aomedia:1373

Change-Id: I388b9cf36817d10613cd2c9d0bd8865b43324009