prefetch-duplex GPU/GPU may be able to associate both streams with a single device

[cwpearson]

If so, we would not need to measure the cost of cudaStreamSynchronize()