Put in a stronger heuristic for the initial plan seed

[haz]

The forest-new benchmark (problems p_10_7 and p_10_10) have initial determinized problems that are too hard for hadd / hff. This is likely also present in the sysadmin probabilistic benchmark.

[haz]

Another option here would be to (1) optionally take an initial plan as a seed; and (2) call a strong classical planner on a determinized version of the initial state and then use (1). Would also open the door to an investigation on the impact having a good first seed plan actually is.