[TensorRT EP] Properly set CUDA_INCLUDE_DIR for onnx-tensorrt (#18274)

#17468
The above PR didn't fully fix the issue for some environments.
This PR fixes this.

cmake/onnxruntime_providers_tensorrt.cmake

@@ -57,12 +57,14 @@
       URL ${DEP_URL_onnx_tensorrt}
       URL_HASH SHA1=${DEP_SHA1_onnx_tensorrt}
     )
+    if (NOT CUDA_INCLUDE_DIR)
+      set(CUDA_INCLUDE_DIR ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}) # onnx-tensorrt repo needs this variable to build
+    endif()
     # The onnx_tensorrt repo contains a test program, getSupportedAPITest, which doesn't support Windows. It uses
     # unistd.h. So we must exclude it from our build. onnxruntime_fetchcontent_makeavailable is for the purpose.
     onnxruntime_fetchcontent_makeavailable(onnx_tensorrt)
     include_directories(${onnx_tensorrt_SOURCE_DIR})
     set(CMAKE_CXX_FLAGS ${OLD_CMAKE_CXX_FLAGS})
-    set(CUDA_INCLUDE_DIR ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}) # onnx-tensorrt repo needs this variable to build
     if ( CMAKE_COMPILER_IS_GNUCC )
       set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} -Wno-unused-parameter")
     endif()