notification: Fix memory leak of pending substreams (#296)

This PR fixes a subtle memory leak that can happen in the following
edge-case situation:
- connection is established and substream outbound is initiated with
remote peer
- the substream ID is tracked until the substream either completes
successfully or fails
- the connection is closed soon after, leading to no substream events
ever being generated

For this edge-cases, we need to remove the tracking of the substream ID
when the connection is reported as closed.

This has been detected after running a node for more than 2 days with
the following generic metrics PR:
- #294

Closes: #295

cc @paritytech/networking

---------

Signed-off-by: Alexandru Vasile <[email protected]>

Author: lexnv

src/protocol/notification/mod.rs

@@ -258,7 +258,7 @@ pub(crate) struct NotificationProtocol {
     /// Connected peers.
     peers: HashMap<PeerId, PeerContext>,
 
-    /// Pending outboudn substreams.
+    /// Pending outbound substreams.
     pending_outbound: HashMap<SubstreamId, PeerId>,
 
     /// Handshaking service which reads and writes the handshakes to inbound
@@ -384,6 +384,8 @@ impl NotificationProtocol {
     async fn on_connection_closed(&mut self, peer: PeerId) -> crate::Result<()> {
         tracing::trace!(target: LOG_TARGET, ?peer, protocol = %self.protocol, "connection closed");
 
+        self.pending_outbound.retain(|_, p| p != &peer);
+
         let Some(context) = self.peers.remove(&peer) else {
             tracing::error!(
                 target: LOG_TARGET,