[BUG] Fixed async event reporting and mutex lock ordering for groups

[ethouris]

Fixes #1631, #1625, #1605, #1603, #1612, #1599, #1585

This PR introduces a common function that is run in case of a broken connection. It takes care of correct error reporting to the application using asynchronous connection methods and sets properly epoll readiness flags for the group.

[maxsharabayko]

• TEST(Bonding, DISABLED_NonBlockingGroupConnect) is still disabled, please enable it as well.

[maxsharabayko]

Changes to srt-test-live should go in a separate PR. This one is already big enough and each review iteration consumes time on looking through changes to apps.
For now I will concentrate on changes to the library.

[maxsharabayko]

Might be worth using debug assertions here and in other places.

SRT_ASSERT(m_GlobControlLock.try_lock() == false);

[ethouris]

Could be, but I'd personally prefer this to be used only in a special thread-tracking mode, in which the lock state would be recorded in a special variable, not just in the debug mode.

[maxsharabayko]

TODO

• Improve connect callback unit test

Take TEST(Bonding, NonBlockingGroupConnect) and modify the ConnectCallback as presented further, and the test will hang up with the current version 3fa5b49 and previous versions as well.

void ConnectCallback(void* /*opaq*/, SRTSOCKET sock, int error, const sockaddr* /*peer*/, int token)
{
    std::cout << "Connect callback. Socket: " << sock
        << ", error: " << error
        << ", token: " << token << '\n';

    srt_close(sock);
}

[maxsharabayko]

(For future reference)

Data race acessing m_iTsbPdDelay_ms

Thread A:
  Write of size 4 at 0x7ba80000c134 by thread T6 (mutexes: write M365):
    #0 CUDT::fillSrtHandshake_HSREQ(unsigned int*, unsigned long, int) core.cpp:1633 (test-srt:x86_64+0x10025e4f2)
	   [NO LOCKS, writes to m_iTsbPdDelay_ms]
    #1 CUDT::fillSrtHandshake(unsigned int*, unsigned long, int, int) core.cpp:1610 (test-srt:x86_64+0x10025e141)
	   [NO LOCKS]
    #2 CUDT::createSrtHandshake(int, int, unsigned int const*, unsigned long, CPacket&, CHandShake&) core.cpp:2231 (test-srt:x86_64+0x10026386c)
	   [NO LOCK, but scope-locks m_parent->m_ControlLock and s_UDTUnited.m_GlobControlLock a bit later to fill HS group extension.
    #3 CUDT::processAsyncConnectRequest(EReadStatus, EConnectStatus, CPacket const&, sockaddr_any const&) core.cpp:4287 (test-srt:x86_64+0x1002872bd)
	   [NO LOCK]
    #4 CRendezvousQueue::updateConnStatus(EReadStatus, EConnectStatus, CPacket const&) queue.cpp:1049 (test-srt:x86_64+0x1003ae9e5)
	   [UNDER m_RIDVectorLock LOCK, can also change m_pUDT->m_bConnecting = false
    #5 CRcvQueue::worker(void*) queue.cpp:1295 (test-srt:x86_64+0x1003b2447)
	   [NO LOCK]

Thread B:
  Previous read of size 4 at 0x7ba80000c134 by main thread (mutexes: write M317):
    #0 CUDTGroup::syncWithSocket(CUDT const&, HandshakeSide) group.cpp:936 (test-srt:x86_64+0x1003f6e6d)
    #1 CUDTUnited::groupConnect(CUDTGroup*, SRT_GroupMemberConfig_*, int) api.cpp:1584 (test-srt:x86_64+0x1001d0e40)
    #2 CUDTUnited::singleMemberConnect(CUDTGroup*, SRT_GroupMemberConfig_*) api.cpp:1284 (test-srt:x86_64+0x1001ccc44)
    #3 CUDTUnited::connect(int, sockaddr const*, int, int) api.cpp:1270 (test-srt:x86_64+0x1001cd5ad)
    #4 CUDT::connect(int, sockaddr const*, int, int) api.cpp:3486 (test-srt:x86_64+0x1001e2cc5)
    #5 srt_connect srt_c_api.cpp:83 (test-srt:x86_64+0x1003dc2e5)

[maxsharabayko]

TODO:

• Update documentation of the connection callback: now only called on failure.

[ethouris]

As per: Data race acessing m_iTsbPdDelay_ms

I think this shouldn't be a problem because makeMePeerOf is a function executed exclusively on the listener side, while groupConnect happens on the caller side. Groups also cannot involve both connecting and accepted sockets. This procedure of setting the latency is also executed exclusively on sockets that are newly being connected and borrow the value from a socket that is already connected.