Predict based on the results of p16x16 search, not final MVs.
This lets us get predictions even if mode decision chose intra.
Also improves cache coherency.