fixing memory ordering issue -- clipping -- adding ncclGetLastError

MSCCL with CUDA graph support

well optimized MSCCL interpreter with NCCL 2.8.4

0.7.2 MSCCL 2.12.12 with CUDA graphs and reduced compilation time

MSCCL 2.12 with CUDA graph support

MSCCL with NCCL 2.12

fully capable MSCCL runtime

increasing the limit for ar_ll128

minor bug fix for how scratchpad is allocated

MSCCL optimized allreduce for up-to 256KB with 8xA100