Performance difference between pytorch direct and forpy coupling

Our hypothesis was that direct coupling would always be faster, but the latest benchmarks show that a forpy based coupling for PyTorch is faster (even if a bit more inconvenient). We are not sure why yet but think it might be something to do with the special support for ndarray's provided by forpy. This needs some more exploration.