On the Importance of Feature Decorrelation for Unsupervised Representation Learning in Reinforcement Learning