GEMM_RS_OVERLAP's sporadic precision issue

[llying-001]

The GEMM reduce-scatter overlap method from Triton Distributed, integrated into this codebase, exhibits intermittent result errors on a specific machine (gpu-44 in the AAC cluster). It runs correctly on other machines. The hardware, ROCm version, Docker image, and the code used are all identical. Even on this machine, the behavior of the unit test case varies at different times.

[llying-001]

maybe related to the code generated by triton, because multi-stream method has not the issue which use the same symm-mem allocation method and reduce kernel.