[BUG] Inner lists in `cuda.bindings` API reference are dropped

Taking NVRTC's --gpu-architecture option as example:

• This is the generated docs (by parsing Doxygen docstrings with an internal tool)

• This is the public docs (the list is too long, so I only show a portion of it)

• This is the docstring from nvrtc.h (also truncated):

 * The valid compiler options are:
 *
 *   - Compilation targets
 *     - \c --gpu-architecture=\<arch\> (\c -arch)
 * 
 *       Specify the name of the class of GPU architectures for which the
 *       input must be compiled.\n
 *       - Valid <c>\<arch\></c>s:
 *         - \c compute_50
 *         - \c compute_52
 *         - \c compute_53
 *         - \c compute_60
 *         - \c compute_61
 *         - \c compute_62
 *         - \c compute_70
 *         - \c compute_72
 *         - \c compute_75
 *         - \c compute_80
 *         - \c compute_86
 *         - \c compute_87
 *         - \c compute_89

We can see the inner lists are not parsed.

Compare the generated API reference with the public docs and the nested list structure shown in the nvrtc.h docstring. Trace the internal Doxygen parsing tool to find where nested lists are dropped, then verify that the CUDA API reference preserves the inner lists.