How Object Information Improves Skeleton-based Human Action Recognition in Assembly Tasks