swarm: connection management behaviours can lead to inconsistent state

In #4777, we discovered that a "connection management" behaviour like libp2p-connection-limits can lead to inconsistent state if it is not applied "first" in the tree of behaviours.

That is because the handle_ functions take &mut self and thus allow modifying the state of a NetworkBehaviour, potentially pre-loading a handler with state. If a behaviour deeper in the tree ends up rejecting the connection, that state is lost.

Preloading handlers is useful to ensure they don't "instant timeout" based on keep-alive after a connection has been established.

The underlying design issue here is that the composition of handle_ functions leads to an inconsistent view across the state of NetworkBehaviours. By having them take &self, this issue would not occur (unless the user users internal mutability but that we can't guard against).

Somehow, we'd need to find a way to still pass data to the handler but only in case the connection is fully established. Perhaps we should find a way to mitigate the "instant shutdown" in a different way by not starting the Connection task until the ConnectionEstablished event has been dispatched to the behaviour?

Start by reading issue #4777 and tracing the NetworkBehaviour handle_ functions involved in connection management and handler preloading. Determine how composed behaviours can avoid retaining state when a deeper behaviour rejects a connection; done requires an agreed design and corresponding coverage, but no file or test is named here.