Accelerating Metropolis-Hastings with Lightweight Inference Compilation