Task-agnostic Continual Learning with Hybrid Probabilistic Models