[FEA] Execute UDFs that provide a RAPIDS execution path

[jlowe]

Is your feature request related to a problem? Please describe.
Some users may wish to provide a UDF that has both a CPU and a GPU implementation, the latter executing directly on the RAPIDS cudf ColumnVector instances just as the plugin operators do. This has a significant drawback in that the user has to implement the UDF twice, but it should provide the best performance when running with the RAPIDS Accelerator.

Describe the solution you'd like
Ideally it would be nice if the user could simply add an interface to their existing UDF class that provides one or more methods for implementing a RAPIDS execution path for the UDF. The plugin would automatically detect the user's UDF implements the required RAPIDS execution interfaces and allow it to execute on the GPU rather than requiring a transition back to the CPU.

Ideally the implementation should provide:

• user documentation on how to implement UDFs with an additional RAPIDS execution path, implemented by Documentation for RAPIDS-accelerated Hive UDFs #1478
• a working example implementing a UDF leveraging cudf Java APIs (i.e.: does not require adding native code), implemented by Support RAPIDS-accelerated HiveSimpleUDF #1393 and Support RAPIDS-accelerated HiveGenericUDF #1408
• a working example implementing a UDF with a custom GPU kernel showing how to build and package the native code, implemented by Add an example RAPIDS-accelerated Hive UDF using native code #1472

Describe alternatives you've considered
We could implement a completely separate interface, specific to the plugin, for registering UDFs. This is less ideal since it's not as seamless.

[jlowe]

Features have been completed, closing this epic.