Welfare under optimal discretionary policy not computed

At least with dennis_1.mod and Gali_discretion.mod from tests/discretionary_policy/, the welfare (as stored in oo_.planner_objective_value) is a NaN. This is a regression from 4.5.

Unfortunately, the test at the end of Gali_discretion.mod did not catch the problem, because of the specific evaluation rules for NaN. This should also be fixed with some call to isnan().