Modin packs multiple apply calls into a queue to reduce ray overhead. We should try to have this feature in modin-gpu as well.