Minimax Statistical Learning with Wasserstein distances