Characterizing the Latent Space of Molecular Deep Generative Models with Persistent Homology Metrics