Network with cache (to not move memory)

[chiamin]

I hope we can use Network but keep the final memory layout for the input tensors, so that we don't need to move memory in every call. This is useful for example in Lanczos iteration in DMRG. I am not sure what is the best way to do it.