[BUG] Fixed desynch between deleted socket and unsubscribing from epoll (misleading log flood)

[ethouris]

There was sometimes reported a log, sometimes with lots of flooding, which complained about event update on a socket that could not be identified. This was happening due to that a socket, which's ID has been already removed from the dispatch table, was trying to update the readiness state in epoll. Finding an already closed socket was impossible, while it was done as the last chance update so that the application picks up the error. This wasn't possible to be done simply because epoll system could not find the socket anyway.

The change was done so that the socket is immediately removed from the epoll system when it's going to be removed from the dispatch table. Note that epoll cannot operate on nonexistent sockets, so it cannot set it any readiness state if it's already closed. Hence removal is done before closing. In the place where the controversial update was done the removal was repeated, just in case, although the socket should be at that moment unsubscribed from any epoll.