Score-based generative models are provably robust: an uncertainty quantification perspective