How to Debug NCCL Performance Issues for ML Workloads?