The experiment results are in:
https://ai.dmi.unibas.ch/_tmp_files/seipp/issue903-v1-issue903-base-issue903-v1-compare.html

I see no noteworthy differences, so I'll go ahead and merge this.

Code looks good. If I understand correctly, the classes AbstractState and
CartesianSet behave differently (it is not just a renaming), in which case a
small experiment might make sense.

I made a pull request at
https://bitbucket.org/jendrikseipp/downward/pull-requests/128 . Could one of you
have a look at the code, please?

Thomas suggested we use CartesianSet instead of AbstractState for temporary
Cartesian sets such as the Cartesian set of goal states. I think this makes a
lot of sense and should make the code easier to understand.