Fix race condition in ExternalThreadEvents

A race condition can occur when the processing thread tries to notify
condition that is not yet waited upon. This can happen during the short
window in the web request thread between transitioning to the WAITING
state and calling condition.wait().

Fix this by locking mutex right before notifying the condition. If the
web request thread is in WAITING state, it is guaranteed to hold the
mutex or wait for condition variable. If the processing thread was able
to transition away from the WAITING state and grabbed the lock, the web
request thread must be waiting on condition variable, so it is safe to
notify it.

Thanks @andrii for tracking the bug down to ExternalThreadEvents and
providing a workaround (grabbing the mutex few lines up than this diff).

hphp/runtime/ext/asio/asio_session.cpp

@@ -111,6 +111,7 @@ void AsioSession::enqueueExternalThreadEvent(c_ExternalThreadEventWaitHandle* wa
     wait_handle->setNextToProcess(nullptr);
     if (m_readyExternalThreadEvents.compare_exchange_weak(next, wait_handle)) {
       // succeeded, notify condition
+      std::unique_lock<std::mutex> lock(m_readyExternalThreadEventsMutex);
       m_readyExternalThreadEventsCondition.notify_one();
       return;
     }