Update base for Update on "add module level benchmark for gemma3 model"

This diff adds a module-level benchmark for the GEMMA3 model.
Also introduce mutlmodal_benchmark.cpp to replace original voxtral_runner.cpp for benchmarking both gemma3 and voxtral model in module level.

Differential Revision: [D84958564](https://our.internmc.facebook.com/intern/diff/D84958564/)

[ghstack-poisoned]

Author: Gasoonjia

extension/tensor/CMakeLists.txt

@@ -24,11 +24,6 @@ target_include_directories(
 )
 target_compile_options(extension_tensor PUBLIC ${_common_compile_options})
 
-# Define USE_CUDA_BACKEND when building with CUDA backend support
-if(EXECUTORCH_BUILD_CUDA)
-  target_compile_definitions(extension_tensor PUBLIC USE_CUDA_BACKEND)
-endif()
-
 # Install libraries
 install(
   TARGETS extension_tensor

extension/tensor/targets.bzl

@@ -10,10 +10,6 @@ def define_common_targets():
     for aten_mode in get_aten_mode_options():
         aten_suffix = ("_aten" if aten_mode else "")
 
-        # Check if USE_CUDA_BACKEND flag is set via build config
-        use_cuda_backend = native.read_config("executorch", "use_cuda_backend", "false") == "true"
-        preprocessor_flags = ["-DUSE_CUDA_BACKEND"] if use_cuda_backend else []
-
         runtime.cxx_library(
             name = "tensor" + aten_suffix,
             srcs = [
@@ -29,7 +25,6 @@ def define_common_targets():
             visibility = [
                 "@EXECUTORCH_CLIENTS",
             ],
-            preprocessor_flags = preprocessor_flags,
             deps = [
                 "//executorch/runtime/core/exec_aten/util:dim_order_util" + aten_suffix,
                 "//executorch/runtime/core/exec_aten/util:tensor_util" + aten_suffix,