Subgaussian and Differentiable Importance Sampling for Off-Policy Evaluation and Learning