FindCUDA: Fix regression in separable compilation without cublas

Revert commit v3.11.0-rc1~274^2 (FindCUDA: Add cublas device library to
separable compilation, 2017-11-03).  It breaks targets that do not link
to cublas.  Another solution will be needed to solve the original
problem in a more compatible way.

Fixes: #17965

Modules/FindCUDA.cmake

@@ -1760,7 +1760,7 @@ function(CUDA_LINK_SEPARABLE_COMPILATION_OBJECTS output_file cuda_target options
       add_custom_command(
         OUTPUT ${output_file}
         DEPENDS ${object_files}
-        COMMAND ${CUDA_NVCC_EXECUTABLE} ${nvcc_flags} -dlink ${object_files} ${CUDA_cublas_device_LIBRARY} -o ${output_file}
+        COMMAND ${CUDA_NVCC_EXECUTABLE} ${nvcc_flags} -dlink ${object_files} -o ${output_file}
         ${flags}
         COMMENT "Building NVCC intermediate link file ${output_file_relative_path}"
         COMMAND_EXPAND_LISTS
@@ -1773,7 +1773,7 @@ function(CUDA_LINK_SEPARABLE_COMPILATION_OBJECTS output_file cuda_target options
         PRE_LINK
         COMMAND ${CMAKE_COMMAND} -E echo "Building NVCC intermediate link file ${output_file_relative_path}"
         COMMAND ${CMAKE_COMMAND} -E make_directory "${output_file_dir}"
-        COMMAND ${CUDA_NVCC_EXECUTABLE} ${nvcc_flags} ${flags} -dlink ${object_files} ${CUDA_cublas_device_LIBRARY} -o "${output_file}"
+        COMMAND ${CUDA_NVCC_EXECUTABLE} ${nvcc_flags} ${flags} -dlink ${object_files} -o "${output_file}"
         COMMAND_EXPAND_LISTS
         ${_verbatim}
         )