Representation Learning for Continuous Action Spaces is Beneficial for Efficient Policy Learning