this is because the JobSystem's queue works as a LIFO, by creating jobs in reverse (memory) order, we attempt to help streaming to the d-cache on that threads -- until the point where jobs are stolen. we also execute the last job immediately instead of creating a job for it -- since we're already in a job.