The Impact of Non-stationarity on Generalisation in Deep Reinforcement Learning