Text this: Contrastive Mask Learning for Self-Supervised 3D Skeleton-Based Action Recognition