Spectral Normalisation for Deep Reinforcement Learning: an Optimisation Perspective