Disentangled Inference for GANs With Latently Invertible Autoencoder