Rebuild pending payments list before replaying pending claims/fails

On `ChannelManager` reload we rebuild the pending outbound payments
list by looking for any missing payments in `ChannelMonitor`s.
However, in the same loop over `ChannelMonitor`s, we also re-claim
any pending payments which we see we have a payment preimage for.

If we send an MPP payment across different chanels, the result may
be that we'll iterate the loop, and in each iteration add a
pending payment with only one known path, then claim/fail it and
remove the pending apyment (at least for the claim case). This may
result in spurious extra events, or even both a `PaymentFailed` and
`PaymentSent` event on startup for the same payment.

Author: TheBlueMatt

lightning/src/ln/channelmanager.rs

@@ -16386,6 +16386,10 @@ where
 			// payments which are still in-flight via their on-chain state.
 			// We only rebuild the pending payments map if we were most recently serialized by
 			// 0.0.102+
+			//
+			// First we rebuild the pending payments, and only once we do so we go through and
+			// re-claim and re-fail pending payments. This avoids edge-cases around MPP payments
+			// resulting in redundant actions.
 			for (channel_id, monitor) in args.channel_monitors.iter() {
 				let mut is_channel_closed = false;
 				let counterparty_node_id = monitor.get_counterparty_node_id();
@@ -16424,6 +16428,18 @@ where
 							);
 						}
 					}
+				}
+			}
+			for (channel_id, monitor) in args.channel_monitors.iter() {
+				let mut is_channel_closed = false;
+				let counterparty_node_id = monitor.get_counterparty_node_id();
+				if let Some(peer_state_mtx) = per_peer_state.get(&counterparty_node_id) {
+					let mut peer_state_lock = peer_state_mtx.lock().unwrap();
+					let peer_state = &mut *peer_state_lock;
+					is_channel_closed = !peer_state.channel_by_id.contains_key(channel_id);
+				}
+
+				if is_channel_closed {
 					for (htlc_source, (htlc, preimage_opt)) in
 						monitor.get_all_current_outbound_htlcs()
 					{