The parallelization functions now keep track of all visited nodes, and do
not visit any node twice anymore. Hopefully, this finally makes
parallelize_mem polynomial in time.