Erez changed the implementation, and it is now much faster. We still need to do
proper timing, but since we need to do that for all heuristics, that's a
separate issue.

The FDTech implementation of reached landmark storage is slower than LAMA's.
We compared using the lama heuristic (without preferred operators) using GBFS,
and got a slowdown of 2-3 times on logistics-00/14-1 (for example).

The implementation should be changed to contain a hash map from states to set of
reached landmarks. This will be done after Silvia integrates the different
landmark/LAMA branches.