Some way to specify consistent precision?

[jacobpennington]

scipy.stats distributions use np.float64 by default, but user-provided values might be anything and it's not clear what precision scipy.minimize (or tensorflow) returns for different input precisions (need to test this). Should make sure consistent precision is kept for all parameters before and after model fits. May also want to give users the option to specify the precision they want to use (i.e. float64 might be unnecessary and wasteful).