Maximizing GPU Utilization Over Multi-Cluster: Challenges and Solutions for Cloud-Native AI Platform