1. Train and init cdfs directly.
2. Use laplace prob estimates in aom_entropy_optimizer to avoid zero
probabilities.

Change-Id: I878fc0a306cbffe3eb51c5b86d5872459b6705c5