On the theoretical foundations of stochastic reduced basis methods