optimize away some device-to-host and host-to-device data copies

[yitao-li]

At the momet there are some copy of data when passing a trained cuML models back to the R process from the GPU. Some of these copies may be reasonably avoided, or at least deferred until data is actually needed on the host.