Inferring Personalized Bayesian Embeddings for Learning from Heterogeneous Demonstration