client should attempt cuda operations early to cause early failure if there's a problem

[arilotter]

we see cuda initialization warnings sometimes, and the process doesnt' crash til we actually train something. we should set up a tiny thing that must execute on GPU to cause early failure, and that can run in a second or so, so we can bail instantly instead of waiting for training to start.