Asymptotic Guarantees for Learning Generative Models with the Sliced-Wasserstein Distance