On the optimal use of provisional data in forecasting with dynamic models