Test keeping output of a model on the GPU

[joewallwork]

The multi-GPU example currently puts the input tensors and model on CUDA devices, but has the output tensors on the CPU. We should test keeping the output on the GPU.