Scale-sensitive Dimensions, Uniform Convergence, and Learnability