On Windows the underlying implementation of QLocalSocket behaves
differently, it seems to trigger the QLocalSocket::readyRead() signal
earlier than anticipated (as on Linux).
Prior to this patch this was not handled gracefully.

Kudos to David Faure for actually figuring this out by just looking at
the code and my debug output on Windows. :)

BEFORE:
Let's have a look at what happens on Windows inside Akonadi's Conneciton
class:

• Akonadi data sends Hello cmd => 91 bytes
• Akonadi resource receives data
• QLocalSocket::readyReady() signal fires
• Triggers Connection::handleIncomingData via signal-slot connection
• On Windows, QLocalSocket::bytesAvailable() is just 8 bytes
• Only the first 8 bytes are read, into the qint64 tag variable
• Protocol::deserialize(...) is called
  • Internally calls waitForData(...) (=> waitForReadyRead()) a few times
• QLocalSocket::readyRead() signal fires again
• Connection::handleIncomingData re-entered => BUG

Problematic end result:

[10972] org.kde.pim.akonadicore: tag: -1099511627647
[10972] org.kde.pim.akonadicore: Invalid command, the world is going to end!
[10972] org.kde.pim.akonadicore: State changed: QLocalSocket::ClosingState
[10972] org.kde.pim.akonadicore: State changed: QLocalSocket::UnconnectedState

> Resource attempts reconnection to server, just to fail again

afterwards.

AFTER:
The fix is to temporarily disconnect from the readyRead() signal while
attempting to wait for data to deserialize commands. This in order to never
re-enter Connection::handleIncomingData() while doing so.