On the Road With 16 Neurons: Towards Interpretable and Manipulable Latent Representations for Visual Predictions in Driving Scenarios