[BENCHMARK][GEMM] Fix poor performance when running CUTLASS Benchmark on BMG (#4338)

This PR fixes the CUTLASS benchmark, which was running poorly on BMG.

It modifies the pinned commit to point to CUTLASS release 3.9.2 and also
adds the missing IGC arguments, as mentioned
[here](https://github.com/codeplaysoftware/cutlass-sycl/blob/sycl-develop/media/docs/cpp/build/building_with_sycl_support.md)

BMG CI:
https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15305727000/job/43057817132

---------

Signed-off-by: Jefferson Le Quellec <[email protected]>

Author: jle-quel

benchmarks/cutlass_kernel/CMakeLists.txt

@@ -4,6 +4,7 @@ set(CUTLASS_KERNEL_FLAGS ${CUTLASS_KERNEL_FLAGS}
   -fsycl-targets=intel_gpu_pvc,intel_gpu_bmg_g21
   "SHELL:-Xspirv-translator=intel_gpu_pvc --spirv-ext=+SPV_INTEL_split_barrier"
   "SHELL:-Xspirv-translator=intel_gpu_bmg_g21 --spirv-ext=+SPV_INTEL_split_barrier"
+  -Xs "-options \"-igc_opts 'VISAOptions=-perfmodel,VectorAliasBBThreshold=1000,ExtraOCLOptions=-cl-intel-256-GRF-per-thread'\" -options -ze-opt-large-register-file"
 )
 
 Python3_add_library(cutlass_kernel MODULE WITH_SOABI python_main.cpp)

benchmarks/cutlass_kernel/cutlass-library.conf

@@ -1 +1 @@
-bb48e86d2fe7cb09eab2e719e78d5811d3da3131
+dee33709bdc0cc579df49f251da894d4546b2624