Speculative pre-positioning: off-path decode for stateful inference sessions

With native support for sessions in an inference engine we can make use of idle GPUs... doing work!