Replaced internal variational refinement implementation inside DeepFlow
with a new optimized one, which provides about the same quality while
being 4-6 times faster.