Device extension should use the CUDA driver library (not the CUDA runtime library)

The initial implementation used the runtime library; however, the driver library is more appropriate (better stability guarantees) plus this is also what Arrow C++ uses in its CUDA implementation.

References:

- https://github.com/apache/arrow-nanoarrow/pull/205#discussion_r1235520762
- https://github.com/apache/arrow-nanoarrow/pull/205#discussion_r1224784183