InteL-VAEs: Adding Inductive Biases to Variational Auto-Encoders via Intermediary Latents