Disconnect nodes in two steps in the full node

[tomaka]

The full node currently can't really sync because the following happens repeatedly: we connect to a peer, start a request, then the peer disconnects us, and we cancel the request.

Why do peers disconnect us? Because they're actually full. Substrate unfortunately doesn't implement the networking protocol as was intended, and instead of refusing peers ahead of time if it is full, it accepts them then disconnects them. See paritytech/polkadot-sdk#556.

Due to the easy-to-use-but-prone-to-race-conditions API, it is not possible to know whether the request might already have a response at the time of the disconnect. If it already has a response, then canceling the request simply discards it, which is meh.
Also, peers might still answer our request even after disconnecting us. It is at their discretion. In fact, Substrate does this. "Disconnecting" in this context doesn't mean "closing the TCP connection" but "notifying that we don't want peering anymore". The TCP connection is only closed after 10 seconds of inactivity.

What this PR does is: when a peer disconnects, we don't cancel the requests that targeting this peer and that are still in progress. Instead, we simply stop sending any new request, and clean up that peer only once all of its requests are finished.

[mergify]

Automatically approving tomaka's pull requests. This auto-approval will be removed once more maintainers are active.

[github-actions]

twiggy diff report

Difference in .wasm size before and after this pull request.

---

 Delta Bytes │ Item
─────────────┼──────────────────
          +0 ┊ Σ [0 Total Rows]