[FEA] Mark all kernels with hidden visibility

[ttnghia]

To avoid potential issues when using a static cuda runtime in the library, we need to hide all kernels from the external world. This can be done by using the static keyword or hidden visibility adding to the kernel declaration, similar to what has been done in cudf: rapidsai/cudf#14726.

For example, we can use this macro instead of __global__:

#ifndef JNI_KERNEL
# define JNI_KERNEL __attribute__ ((visibility ("hidden"))) __global__
#endif

[mattahrens]

Additional documentation needed to show how to do this properly with UDFs.

[jlowe]

Additional documentation needed to show how to do this properly with UDFs.

Filed NVIDIA/spark-rapids-examples#356 to track updating the native UDF examples in the spark-rapids-examples repository.