routing: undirected node pair based mission control [wip] [don't review]

[joostjager]

Builds onto #1734

Previously mission control tracked failures on a directed node pair basis. This PR changes this to tracking on the level of undirected node pairs. The goal of moving to this coarser-grained level is to reduce the number of required payment attempts without compromising payment reliability.

Deeper analysis of channel-level failures and the way they should be attributed to nodes suggests that there always remains an uncertainty about which of the two endpoints is responsible for the failure. It is therefore appropriate to penalize the pair of nodes as a whole. This penalizes the bad node, while the good node only sees its connection to the bad node penalized. This prevents bad node pairs from being tried in both directions.

An exception is insufficient channel balance (FailTemporaryChannelFailure). Undirected penalization may seem too aggressive in this case, but:

• It is likely that channels are used for sending payments in a single direction only, away from the sender. Theoretically it is possible to connect to the graph in such a way that a channel may get used both ways, but this seems rare. (Note: this is just about sending. Receiving obviously utilizes channels in the other direction, but for that, path finding is the other party's responsibility.)
• One could argue that a channel that isn't balanced in one direction, should be avoided in the opposite direction too. It could be a sign that one of the nodes is not properly taking care of the channel (and possibly its other channels).
• If we are more tolerant to temp. chan. failures, nodes may use that to downgrade their penalty. Example: a node advertises an unreliable public channel. The next peer is offline. By sending back a temp. chan. failure instead of an unknown next peer failure, the peer that is offline isn't penalized. This may be good for the failure-sending node because it may generate incoming traffic in the future.

This PR also prepares for #3256 where a correct, gaming-resistent, set of failure attribution rules is implemented.

[joostjager]

Closing this pr for now. I still think we will eventually be forced to move towards undirected pair-based penalization to be fully gaming resistent, but this will probably not be on the short term. Will keep this pr around for when that moment comes.