msccl-v2.19

1. catch up with nccl 2.19.4
2. add alltoall support
3. optimized for A100/H100

msccl-v2.18

1. Catch up the nccl 2.18
2. Improved capability for lower level of CUDA
3. optimized algo file included so just download to use without any extra optimization steps in common scenarios.

msccl-v2.17

New Features

• Sync up to NCCL 2.17.
• Support unified MSCCL executor interface.
• Support FP8 data type.
• Enhanced multi-node performance with CUDA graph enabled.