fix(online): Make P2P connections and the lobby flow reliable

[tintinhamans]

Since switching to native ICE, more players got "cannot connect to all players". GNS now ships with both WebRTC and native ICE, selected by ice_implementation in the service config (default WebRTC, which worked before). The native client now pairs relay candidates found late and tries all STUN/TURN servers at once, so testers can be moved to it without a client release.

• The cannot-connect handler no longer deletes the mesh during its own update, which crashed the game.
• Only the later joiner of a failing pair gives up (2 attempts). A departed peer is dropped without leaving the lobby.
• Native ICE also ranks network adapters, uses one stable IPv6 address per adapter, requests IPv6 relays, gathers again when the network changes, and fails over to a working backup pair when the active one goes quiet.
• GNS starts once per login instead of every lobby. Mesh init failures fail cleanly. Anticheat connection-state updates are applied on the main thread.
• Websocket messages and signals are queued instead of dropped, and flushed before disconnect. The host's connectivity check always resolves.
• Lobby: no crash on Accept after the lobby is gone, stale create/join replies ignored, 64-bit lobby IDs kept, slots set by the server's slot index.
• Quick match: callbacks ignored after the menu closes, failures decided by the matchmaker, setup timeout limited, quick match flag reset.