Identifiable Deep Generative Models via Sparse Decoding