Add GEMM-based standard SDPA benchmark

Add bench_sdpa.cpp with a standalone GEMM-based SDPA implementation
(run_standard_sdpa) alongside ExecuTorch's tiled flash attention
(custom_sdpa_out) for comparative benchmarking.

The standalone SDPA uses full GEMM per head with 3-pass softmax and
supports both [B,S,H,D] and [B,H,S,D] layouts via BLAS leading
dimension parameters, allowing isolation of algorithm vs layout effects.

Includes validation tests that verify the GEMM-based implementation
matches custom_sdpa_out within tolerance.

Differential Revision: [D96044313](https://our.internmc.facebook.com/intern/diff/D96044313/)

ghstack-source-id: 361224784
Pull Request resolved: #18646

Author: kimishpatel

extension/llm/custom_ops/BUCK

@@ -1,4 +1,5 @@
 load("@fbcode_macros//build_defs:build_file_migration.bzl", "fbcode_target", "non_fbcode_target")
+load("@fbcode_macros//build_defs:cpp_benchmark.bzl", "cpp_benchmark")
 oncall("executorch")
 # Any targets that should be shared between fbcode and xplat must be defined in
 # targets.bzl. This file can contain xplat-only targets.
@@ -89,3 +90,16 @@ fbcode_target(_kind = runtime.python_test,
         "//caffe2:torch",
     ],
 )
+
+fbcode_target(_kind = cpp_benchmark,
+    name = "bench_sdpa",
+    srcs = ["bench_sdpa.cpp"],
+    deps = [
+        "fbsource//third-party/benchmark:benchmark",
+        "//executorch/extension/llm/custom_ops:custom_ops_mkl_noomp",
+        "//executorch/extension/threadpool:threadpool",
+        "//executorch/kernels/optimized:libblas",
+        "//executorch/runtime/core/exec_aten:lib",
+        "//executorch/runtime/core/exec_aten/testing_util:tensor_util",
+    ],
+)