On the optimality of kernels for high-dimensional clustering