using avx to accelerate inter node protoll

[visualxu]

hi.
I conducted some experiments using the avx instruction on inter_node_proto_ll and achieved better performance.
Performance with avx/without avx:

Here is the environment:
Intel(R) Xeon(R) Silver 4316 CPU @ 2.30GHz
256GB DRAM
A100 80GB * 2
100gbps mellaonx5
This is a reference for simple implementation(buffer address/size must be aligned with 256/512):
visualxu@1c502b2#diff-45a9034a0c75cbfbbb34e853a43f6513c1d4c933eccf6adca705abe234fc1113
I hope this idea can be helpful to the NCCL community.