the error in reconstructing the k-step ahead forecasts with occbin smoother...

the error in reconstructing the k-step ahead forecasts with occbin smoother redux, should only occur in the first occbin smoother interation for non-stationary models (with diffuse filter), where tstart>1.

if k>tstart, the same simulation should have been done already in occbin.kalman_update, so it should never give an error

if k<=tstart, the simulation may crash, since we ignore OBC in the first (diffuse) steps and it may happen that, given the linear updated states, the occbin simulation does not converge.

Indeed, I never got a simulation failure in line 474, where this possibility is properly taken into account.

(cherry picked from commit abbd1118)