Optimal prior-dependent neural population codes under shared input noise