[core] Fixed problem with handling the lingering links case in non-blocking group reader

[ethouris]

The problem is described in #1805, though may need further testing.

This PR also contains a fix for lock ordering at the handshake, might need to be moved.

[ethouris]

This remains as draft due to enclosing two different solutions. The fix for wrong lock on m_ControlLock was necessary for proper testing the problem. The fix for wrong epoll update has been tested by the issue reporter.

[gou4shi1]

why not merge this pr?

[ethouris]

I remerged this with latest upstream; the fix for the deadlock on ConnectionLock is already merged. The remaining thing is the avoidance of the spurious read-ready for group.

[ethouris]

Note that I have doubts about the provided solution mainly because this newly introduced flag seems to be required to be filled from a separate thread, so likely this field should be atomic. There's another PR that provides atomic and should be cleared first.

[gou4shi1]

Great, I found your atomic PR is merged, so this PR is unblocked now?

The remaining thing is the avoidance of the spurious read-ready for group.

The remaining part I care is just if (ready_sockets.empty())) throw CUDTException(); which should be a hotfix, LOL

[maxsharabayko]

Closed by #2203