Support sink_parquet/sink_ndjson/sink_csv with GPU engine

[wence-]

Description

When collecting a lazy query with the GPU engine, we can currently only materialise the end result in CPU memory. If we want to end up with the final result "at rest" as parquet (or other on-disk format), we incur the cost of moving to CPU memory and then writing from there. libcudf has GPU-accelerated IO writers as well as readers, so it would be nice to be able to sink straight from GPU memory.

I think this is doable in the same way we currently hook into collect, by adding an engine argument.