Neural Wasserstein Gradient Flows for Maximum Mean Discrepancies with Riesz Kernels