better handling of move_to_first

[CJ-Wright]

Current behavior causes the graph to be recreated as it was initialized, if move_to_first is used then we don't capture the changes to the graph structure. We need to do that better.
This is particularly difficult as a move_to_first use case is handling tricky diamond dependencies where we can't actually init downstream things with the correct execution order.

One approach could be to stash the upstreams in the serialized graph. On deserialization we could then create all the nodes via the init information and then finally reorder the graph accordingly.