ci: harden cuda include path handling (#15093)

On windows we can get multiple include dirs, so find where the headers are then
copy from that location.

CMakeLists.txt

@@ -284,7 +284,19 @@ if(MLX_ENGINE)
     # The Go mlxrunner sets CUDA_PATH to OLLAMA_INSTALL_DIR so MLX finds them at
     # $CUDA_PATH/include/*.h via NVRTC --include-path.
     if(CUDAToolkit_FOUND)
-        set(_cuda_inc "${CUDAToolkit_INCLUDE_DIRS}")
+        # CUDAToolkit_INCLUDE_DIRS may be a semicolon-separated list
+        # (e.g. ".../include;.../include/cccl"). Find the entry that
+        # contains the CUDA runtime headers we need.
+        set(_cuda_inc "")
+        foreach(_dir ${CUDAToolkit_INCLUDE_DIRS})
+            if(EXISTS "${_dir}/cuda_runtime_api.h")
+                set(_cuda_inc "${_dir}")
+                break()
+            endif()
+        endforeach()
+        if(NOT _cuda_inc)
+            message(WARNING "Could not find cuda_runtime_api.h in CUDAToolkit_INCLUDE_DIRS: ${CUDAToolkit_INCLUDE_DIRS}")
+        else()
             set(_dst "${OLLAMA_INSTALL_DIR}/include")
             set(_MLX_JIT_CUDA_HEADERS
                 builtin_types.h
@@ -320,6 +332,7 @@ if(MLX_ENGINE)
                 DESTINATION "${_dst}/crt"
                 COMPONENT MLX)
         endif()
+    endif()
 
     # On Windows, explicitly install dl.dll (dlfcn-win32 POSIX dlopen emulation)
     # RUNTIME_DEPENDENCIES auto-excludes it via POST_EXCLUDE_FILES_STRICT because