LAVA: Latent Action Spaces via Variational Auto-encoding for Dialogue Policy Optimization