Feature Likelihood Score: Evaluating the Generalization of Generative Models Using Samples