Deal with treatment of presample

#429 introduces an error messsage for non-linear estimation with a presample being specified. This incorrectly used the first presample observations to compute the likelihood despite the manual saying that they are skipped. In the long-run, we might either want to actually use the presample to initialize the state matrix (e.g. as an option for initialization similar to using the ergodic state variance matrix of the linear model) or just skip the first presample periods by only using the observations presample+1:end

Another place where presample shows up is bvar_toolbox.m. Here we should clarify the distinction to the training sample in options_.bvar_prior_train