Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training