On Russian Roulette Estimates for Bayesian Inference with Doubly-Intractable Likelihoods