Add XeTLA FA backward implementation to benchmark (#2367)

Implementation comes from
[IPEX](https://github.com/intel-innersource/frameworks.ai.pytorch.ipex-gpu/blob/e0026ba2e9e5d3d1601059334b4d013ec77a4625/csrc/gpu/aten/operators/xetla/kernels/SDP/fmha_backward.cpp)


https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11089057652

---------

Co-authored-by: Anatoly Myachev <[email protected]>

Author: yudongsi

benchmarks/xetla_kernel/CMakeLists.txt

@@ -2,7 +2,10 @@
 find_package(XeTLALibrary REQUIRED)
 set(CMAKE_CXX_STANDARD 20)
 
-set(XETLA_KERNEL_FLAGS ${XETLA_KERNEL_FLAGS} -fsycl)
+set(XETLA_KERNEL_FLAGS ${XETLA_KERNEL_FLAGS}
+  -fsycl
+  -fsycl-device-code-split=per_kernel
+)
 
 if (USE_AOT_DEVLIST)
   set(XETLA_KERNEL_FLAGS ${XETLA_KERNEL_FLAGS} -fsycl-targets=spir64_gen)