Dont create duplicate nodes in graph merges

[vadixidav]

deep/deep/src/lib.rs

Lines 79 to 85 in 7cc85ca

	fn merge(&mut self, other: Graph) {
	let current = self.ops.len();
	self.ops.extend(other.ops);
	for op in &mut self.ops[current..] {
	op.shift_inputs(current);
	}
	}

In this code, if a graph merges twice (the same tensor is used in two places) then it will create two separate trainable graphs where each one only gets the gradient from one of the places its used. To avoid this, any time a graph is modified, we should create a random unique identifier of every range of ops that was previously part of an older graph. These ranges should be preserved when merging graphs as well. If the unique identifier from a range is found in a graph that is being merged, we should reuse that part of the graph by adding its starting offset minus the starting offset in the other graph. This will allow us to create arbitrary DAGs (directed acyclic graphs) without issue.

This wont surface as an issue if a piece of graph is never used in two different places, which is actually surprisingly rare. This issue is just to track this problem to avoid it randomly appearing unexpectedly.