VECO 2.0: Cross-lingual Language Model Pre-training with Multi-granularity Contrastive Learning