Dialing a node for which we have no address doesn't notify the PSM

[tomaka]

I have some sub-libp2p logs that show a buggy situation:

2020-04-17 11:42:03.441 tokio-runtime-worker DEBUG sub-libp2p  PSM => Connect(PeerId("QmWabGpJf7JyezL2bpsykobTMecrEDeT21d4W6ntcv39Zx")): Starting to connect
2020-04-17 11:42:03.441 tokio-runtime-worker DEBUG sub-libp2p  Libp2p <= Dial PeerId("QmWabGpJf7JyezL2bpsykobTMecrEDeT21d4W6ntcv39Zx")
2020-04-17 11:42:03.441 tokio-runtime-worker TRACE sub-libp2p  Addresses of PeerId("QmWabGpJf7JyezL2bpsykobTMecrEDeT21d4W6ntcv39Zx") are []
2020-04-17 11:42:03.441 tokio-runtime-worker DEBUG sub-libp2p  Requested dialing to PeerId("QmWabGpJf7JyezL2bpsykobTMecrEDeT21d4W6ntcv39Zx") (peer not in k-buckets), and no address was found

...and then, importantly, no further logs related to QmWabGpJf7JyezL2bpsykobTMecrEDeT21d4W6ntcv39Zx

(note that this is before #5679, which slightly changed the logs format here)

What happens is:

• The NetworkBehaviour::poll emits a DialPeer event.
• The libp2p Swarm then calls addresses_of_peer, but the method returns an empty list of addresses.
• In libp2p 0.17, this would then call inject_dial_failure, which would then print additional logs and report to the peerset that the dialing attempt has failed. In libp2p 0.18, however, this doesn't seem to happen anymore.

The consequence is that the peerset and the NetworkBehaviour think that we're connecting to QmWabGpJf7JyezL2bpsykobTMecrEDeT21d4W6ntcv39Zx, while in fact we are not.

cc @romanb Could you look into this? I'm not necessarily familiar with the state machine anymore to know where this should be fixed.
EDIT: Roman is on vacation

[tomaka]

This is likely to be the reason for #5594