Enforce estimation to be fully Bayesian or fully ML

Currently, we allow for mixtures in the estimated_params-block like

estimated_params;
csadjcost,6.3325,2,15,NORMAL_PDF,4,1.5;
csigma,1.2312,0,1;

This leads to strange subsequent crashes like

Error using prior_bounds (line 157)
prior_bounds: unknown distribution shape (index 1, type 0)

We should check earlier whether priors have been defined for all variables or none.