Separation Results between Fixed-Kernel and Feature-Learning Probability Metrics