Weighting the past and the present: Neural computations for probabilistic inference about rewards