* fix cuda toolkit lookup and parallel * support user override first * enable control over the nested parallel count