Temporal-masked skeleton-based action recognition with supervised contrastive learning