Fix/reconnect multinode timeout

[Ente3]

Motivation

When the connection between two nodes dropped, even for a moment, both nodes removed each
other from the cluster immediately and never tried to reconnect. The cluster stayed split
until a node was restarted, and each node continued as its own head node. This happens
more often with nodes in different datacenters and with more nodes in the cluster.

The reconnect grace period (cloudnet.max.node.disconnect.millis) only applied to nodes
that stopped sending updates while their socket stayed open. When the socket itself was
closed, the channel close handlers called NodeServer#close() directly. The reconnect
handshake had also never really been used, and it had several problems once it was.

Modification

• A lost node connection now enters the reconnect grace period when
  cloudnet.max.node.disconnect.millis is greater than 0, instead of removing the node
  directly. The default of 0 keeps the previous behaviour.
• Only one reconnect attempt per node is in flight at a time, because concurrent
  connections can break the reconnect handshake. Nodes stuck in the syncing state after
  an interrupted reconnect are removed once the grace period expires.
• Nodes send their startup time when they authorize. The other node can then detect a
  restart during the grace period and let the node join as a newly started node, instead
  of reconnecting it into its old session (which failed with an NullPointerException in
  AuthorizationResponsePacketListener). Nodes on older versions don't send it, and
  older nodes ignore it, so mixed versions keep working.
• Remote service providers look up the current channel of their node on each call,
  instead of keeping the channel that was used during registration. That channel is
  closed after a reconnect.
• The service sync during a reconnect now applies the reconnecting node's services.
  Before, they were skipped because the node wasn't available() yet.
• Queued packets are flushed when a reconnect ends in a plain authorization, and local
  services are re-published to the other node.
• The head node is re-selected when a node becomes ready again, so all nodes agree on
  the same head node.
• The grace period is read when NodeDisconnectHandler is created, so it can be set in
  tests.
• Added unit tests for NodeDisconnectHandler, NodeDisconnectTrackerTask,
  AuthorizationPacketListener and AuthorizationResponsePacketListener.
• Updated the launcher.cnl description of cloudnet.max.node.disconnect.millis.

Result

With a grace period configured, nodes reconnect after short network interruptions and
agree on the same head node afterwards. A node that restarts within the grace period
joins cleanly as a new node. If a node does not come back within the grace period, it is
removed like before. With the default setting, behaviour is unchanged.

Tested manually with two nodes (grace period 30s/60s):

• Killing the node connection (ss -K): before, both nodes dropped each other and never
  reconnected. Now both are READY again within a few seconds, with the same head node.
• Blocking the connection with a firewall (idle timeout): reconnects, and the head node
  now matches on both nodes (before, each node showed itself as head).
• Crashing and restarting a node within the grace period: the restart is detected and
  the node joins as a new node without errors.
• Outage longer than the grace period: nodes are removed after the grace period as before.
• Grace period 0: same behaviour as before.

Other context

Known limitations:

• If a node does not reconnect within the grace period, the cluster stays split until a
  node is restarted, like before.
• During the grace period, services and players of the disconnected node stay registered.
• The service-related changes in DefaultCloudServiceManager are not covered by unit
  tests, as the class needs most of the node to be constructed.

Tests:
https://hastebin.baxnetzwerk.de/ct-Z4gFrfyk
With cloudnet.max.node.disconnect.millis set to 60.000:
https://hastebin.baxnetzwerk.de/dzvj7KguKXk
With cloudnet.max.node.disconnect.millis set to 0 or default:
https://hastebin.baxnetzwerk.de/Bz+rvVdSEei
(the links are available for 1 year.)

Fixes #1897