`peer_disconnected` event processing isn't robust if a peer_connected Errs

We guarantee that `peer_disconnected` will always be called for a peer after `peer_connected`, but we don't if one message handler fails `peer_connected` (the others wont even get it). Also, `PeerManager::read_event` isn't clear about if the user has to call back into `PeerManager` with a disconnection event (i think so, but we should check ldk-garbagecollected/ldk-c-bindings/ldk-net-tokio).