Off-Policy Evaluation and Learning for External Validity under a Covariate Shift