Partitioning variability in animal behavioral videos using semi-supervised variational autoencoders