On the Identifiability of Hybrid Deep Generative Models: Meta-Learning as a Solution