Consider creating "optimizer state" node type

[tomiock]

These types of node need to be handled as parameters -they are constant during training and should not be reallocated/freed-, however, they do not require a gradient.