Refined Learning Bounds for Kernel and Approximate k -Means