Calibrating Neural Simulation-Based Inference with Differentiable Coverage Probability