Implicit encoding of prior probabilities in optimal neural populations