Emergent Bayesian Behaviour and Optimal Cue Combination in LLMs