Approximate Bayesian inference from noisy likelihoods with Gaussian process emulated MCMC