For SALE: State-Action Representation Learning for Deep Reinforcement Learning