Open-Set Incremental Learning via Bayesian Prototypical Embeddings