Learning with Differentially Private (Sliced) Wasserstein Gradients