[issue383] Generate reasonable landmark orderings deterministically.

Approximating reasonable landmark orderings was non-deterministic prior to this
commit. The change fixes this issue and leads to more reasonable orderings being
generated (1% more with `lm_zg`, 2% more with `seq-opt-bjolp`, 13% more with
`lm_rhw`, and 800% more with `lm_hm(m=2)`). More orderings slow down the
heuristic evaluation but generally improve heuristic accuracy. This leads to
significantly fewer expanded states in some cases of optimal planning and
improved plan quality in satisficing planning.

src/search/landmarks/landmark_factory_reasonable_orders_hps.cc

@@ -50,9 +50,6 @@ void LandmarkFactoryReasonableOrdersHPS::approximate_reasonable_orders(
         if (landmark.disjunctive)
             continue;
 
-        if (landmark.is_true_in_state(initial_state))
-            return;
-
         if (landmark.is_true_in_goal) {
             for (auto &node2_p : lm_graph->get_nodes()) {
                 const Landmark &landmark2 = node2_p->get_landmark();