A Bias-Variance-Covariance Decomposition of Kernel Scores for Generative Models