Done and merged. Actually, there were no further improvements -- I did some code
cleanup, but it had negligible impact on speed. (It may actually be slightly --
very slightly -- negative, but there were good reasons for it.)

While working on the heuristic, I've also removed the iteration_limit option
which was unused and did not make much sense for the cost-based setting since it
was not cost-based.

OK, should be much better now. This is not thoroughly tested; performance in
unit-cost domains might drop a bit. But initial tests at least don't suggest
that this would be the case.

I'm not done yet, as further improvements are possible, but I've already pushed
what I have so far and merged it into default, without closing issue116 (since
I'll work further on this later).

See also issue173.

The queue used for h_max computation in LM-cut is bucket-based and hence works
very poorly with high action costs as in ParcPrinter.