Fix deserialization of similar token networks

[palango]

So that turned out to be quite a bug, much more foundational than I initially thought.

From initial debugging with @CosminNechifor we saw that there was a mismatch in the internal network graph which is part of the TokenNetworkGraphState. However all tries to reproduce this failed, the correct state changes were written to the WAL and when replaying the WAL everything was right.
However, we finally found that we could reproduce this when we started with a snapshot. From here on the problem was easy to track down.

It turns out that that there was a problem during deserialization which made multiple TokenNetworkState objects reference the same TokenNetworkGraphState. This only happened in really special constrains, so it’s not surprising that we haven’t found this earlier:
A snapshot need to be deserialized when there are multiple token networks with the same topology and exactly the same channel participants at the time of the snapshot.
So thanks to the scenario player @CosminNechifor was able to trigger that.

For the fix: It turns out that the TokenNetworkGraphState currently doesn’t know which token network it belongs to. We therefore need a breaking change and add the token network identifier to the TokenNetworkGraphState. With this the objects don’t compare to equal any more and proper instances for each token network are created during deserialization.

This is the easy and quick fix, but in the future we might want to move the TokenNetworkGraphState in the TokenNetworkState or remove it all together when the PFS gets introduced.

Fixes #2662

[rakanalh]

I think the change makes sense. As i understand that the RefCache's logic was caused this as one instance was created instead of multiple expected TokenNetworkGraphState, correct?

[palango]

As i understand that the RefCache's logic was caused this as one instance was created instead of multiple expected TokenNetworkGraphState, correct?

Exactly. This happened easiliy when running the same scenario multiple times with the same nodes. Then the comparison method would return true, even though there should have been different graph states.

raiden/raiden/transfer/state.py

Lines 507 to 520 in 5432b6f

	def __eq__(self, other):
	return (
	isinstance(other, TokenNetworkGraphState) and
	self._to_comparable_graph() == other._to_comparable_graph() and
	self.channel_identifier_to_participants == other.channel_identifier_to_participants
	)
	def __ne__(self, other):
	return not self.__eq__(other)
	def _to_comparable_graph(self):
	return sorted([
	sorted(edge) for edge in self.network.edges()
	])