Optimized primitives for collective multi-GPU communication - NVIDIA/nccl