Imported from https://github.com/socketio/socket.io-redis-adapter/blob/ef5f0da0b4928fd422afc985aec0e233d34400c0/lib/cluster-adapter.ts

This abstract class is currently used by the sharded Redis adapter. We
import it here because the logic can be reused by the other adapters,
which will then only need to handle the pub/sub mechanism.

Related: socketio/socket.io-redis-adapter@bd32763

The clustered adapter will now:

- periodically check if a node has left the cluster
- send an event when it leaves the cluster

This should reduce the number of "timeout reached: only x responses
received out of y" errors.

The Redis adapter is currently the only adapter which makes a
distinction between publishing messages (one channel for all) and
responses (one channel for each node).

Note: the current API does not currently allow the user to handle those
errors (and retry, for example). This might be worth investigating for
the next major version, maybe something like:

```js
try {
  await io.emit("hello");
} catch (e) {
  // something went wrong
}
```

Related: socketio/socket.io-mongo-adapter#15

Reference: https://nodejs.org/api/timers.html#timeoutrefresh

The test suite is imported from [1].

[1]: https://github.com/socketio/socket.io-postgres-adapter/blob/193883b9caefee4c6fa31f49d5f5197f4d9a8186/test/index.ts

Reference: https://github.com/nodejs/Release

Reference: https://github.blog/changelog/2023-09-22-github-actions-transitioning-from-node-16-to-node-20/

The fetchSockets() method of the parent class is synchronous, so the
onMessage() method does not need to be asynchronous.

Diff: 2.5.2...2.5.3