Model-free Posterior Sampling via Learning Rate Randomization