Model-free Representation Learning and Exploration in Low-rank MDPs