[GEMM StreamK] Investigate and optimize performance

[LiyangLingIntel]

Triton GEMM streamk performance still has a gap to XeTLA.

We need to investigate XeTLA implementation and optimize Triton implementation.

[LiyangLingIntel]

With the fix in #3374, the bechmark streamk case can get ~1.4x speedup than the previous one, and it can get ~0.97x than XeTLA, but it's still slower than the common implementation, ~0.85x.

[LiyangLingIntel]

With the fix in #3374, the bechmark streamk case can get ~1.4x speedup than the previous one, and it can get ~0.97x than XeTLA, but it's still slower than the common implementation, ~0.85x.

Update the benchmark perf data on PVC1550, the Triton StreamK can get 70~80% of XeTLA, we still need to do optimization.

The 2 directions:

1. Reduce threads race on index when lowering AtomicOp.
2. Optimize the StreamK triton kernel, refer to XeTLA implemetation and new published work(https://arxiv.org/pdf/2408.11417)