It turns out that you can transfer CUDA memory to the host around 2x
faster if you have pinned memory in the host. That is tricky since
buffers are allocated internally in IceT. Change the render image
fetches to allow them to allocate their own memory.