Correct _LIBCUDACXX_CUDACC_VER value computation (#9579)

This brings the libcudacxx patch inline with the proposed upstream changes ( NVIDIA/libcudacxx#225 ).

This is needed so that CUDA versions are correctly computed and decimal-128 support is enabled when using CUDA 11.5+

Authors:
  - Robert Maynard (https://github.com/robertmaynard)

Approvers:
  - Conor Hoekstra (https://github.com/codereport)
  - Mark Harris (https://github.com/harrism)

URL: #9579

cpp/cmake/libcudacxx.patch

@@ -7,7 +7,7 @@ index d55a43688..654142d7e 100644
      #define _LIBCUDACXX_CUDACC_VER_BUILD __CUDACC_VER_BUILD__
      #define _LIBCUDACXX_CUDACC_VER                                                  \
 -        _LIBCUDACXX_CUDACC_VER_MAJOR * 10000 + _LIBCUDACXX_CUDACC_VER_MINOR * 100 + \
-+        _LIBCUDACXX_CUDACC_VER_MAJOR * 10000 + _LIBCUDACXX_CUDACC_VER_MINOR * 1000 + \
++        _LIBCUDACXX_CUDACC_VER_MAJOR * 100000 + _LIBCUDACXX_CUDACC_VER_MINOR * 1000 + \
          _LIBCUDACXX_CUDACC_VER_BUILD
 
      #define _LIBCUDACXX_HAS_NO_LONG_DOUBLE