[BMG][FlexAttn] Unbelievably high perf with DLE 2025.3

[whitneywhtsang]

https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/19744824959/job/56576886398

    Z  H_q  H_kv  N_CTX_q  N_CTX_kv  D_HEAD_qk  D_HEAD_v MODE  Triton-GB/s  Torch-GB/s  Triton-GB/s-min  Torch-GB/s-min  Triton-GB/s-max  Torch-GB/s-max  Triton-TFlops  Torch-TFlops  Triton-TFlops-min  Torch-TFlops-min  Triton-TFlops-max  Torch-TFlops-max  Triton-CV  Torch-CV
4   1  128     1      512      1664        576       512  fwd   307.910042    9.630185        18.468514        7.563575       309.257742        9.659267     461.752076     14.441744          27.695993         11.342587         463.773132         14.485356   0.001911  0.002321

[whitneywhtsang]

@anmyachev Can you please take an initial look, maybe problem with PTI?

[whitneywhtsang]

Started a CI run with elapsed time: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/19747687726

[whitneywhtsang]

Started a CI run with elapsed time: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/19747687726

Performance is good with elapsed time.