Update on "gemma3 e2e runner on cuda"

This diff introduces e2e runner for gemma3 model on cuda delegating using AOTI library, which is guarded by CI. Also other necessary infrastructure updates for building and running the `gemma3 e2e runner` on CUDA devices.

Differential Revision: [D85087532](https://our.internmc.facebook.com/intern/diff/D85087532/)

[ghstack-poisoned]

Author: Gasoonjia

examples/models/gemma3/CMakeLists.txt

@@ -117,6 +117,8 @@ if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
   endif()
 endif()
 
-target_include_directories(gemma3_e2e_runner PUBLIC ${_common_include_directories})
+target_include_directories(
+  gemma3_e2e_runner PUBLIC ${_common_include_directories}
+)
 target_link_libraries(gemma3_e2e_runner PUBLIC ${link_libraries})
 target_compile_options(gemma3_e2e_runner PUBLIC ${_common_compile_options})

extension/llm/runner/multimodal_prefiller.cpp

@@ -94,7 +94,9 @@ Result<uint64_t> MultimodalPrefiller::prefill(
     if (expected_dtype == ::executorch::aten::ScalarType::BFloat16) {
       // Convert to bfloat16 for model input
       auto image_tensor_return = convert_to_bfloat16(image_tensor);
-      ET_CHECK_OK_OR_RETURN_ERROR(image_tensor_return.error(), "Failed to convert image tensor to bfloat16");
+      ET_CHECK_OK_OR_RETURN_ERROR(
+          image_tensor_return.error(),
+          "Failed to convert image tensor to bfloat16");
       image_tensor = image_tensor_return.get();
     }