Maximum Likelihood Learning of Unnormalized Models for Simulation-Based Inference