In-context reinforcement learning through bayesian fusion of context and value prior