Adversarial learning of neural user simulators for dialogue policy optimisation