Don't flip the image at all: just read in the opposite direction,
and transmit flipped position.

Only convert the images that we are actually going to transmit. TODO:
We should keep the image as non-converted and only convert the actual
pixels that we are sending.

Let the client keep a reference to the current image. This is a
prerequisite for doing RGB swapping and flipping on the fly. In
addition, this should also fix a theoretical race condition.

Bug caused by misreading documentation. The docs say:

"If another socket is created in the reimplementation
 of this method, it needs to be added to the Pending
 Connections mechanism by calling
 addPendingConnection()."

However, this is only if you actually want to use the
Pending Connection mechanism. When the socket is added
to this list, it will be deleted by the QTcpServer on
close, and since it was also deleted by the QVncClient
this would crash.

Although the spec does not say the padding has to
be zero, it looks like RealVNC uses the padding for alpha
if non-zero.

This should make it work on OpenGL ES, but it is very inefficient
since it converts the whole image twice for each screen grab

We put the clients and their sockets into separate
threads, so that communication, compression, dirty
map resolution, etc. is all done out of the main
thread.

...and some cleanups

When the surface is resized (or anything that triggers
SetEncodings), we recreate the encoder, but the zlib
stream should retain its state. So we store this as part
of the client instead.

The specs require that you have a single stream object for the
connection which does not close until the connection closes,
retaining state between frames. Failing to do this caused clients
to error out with Z_STREAM_END, which should never be reached
as long as the condition is met.

qCompress() adds the uncompressed length of the buffer to the
beginning of the stream, which is not part of the protocol.

However, this is still disabled, because while the first frame now
looks fine, subsequent frames get errors.

(name subject to change)

This just builds on raw and compresses the content with
zlib, so it sounds like the easiest compression available

This was probably there for a reason in the original, and the
issue it was uncommented to fix has been fixed in a better way.

Manual test draws text of key events into image now.

It was possible to get mouse press events without mouse release
events, e.g. if one client sent a Qt::LeftButton press and another
client sent a Qt::RightButton press. The Qt::RightButton would
just clear the right button state without sending any event for it.

Q_DECLARE_CLASS declares it as class, so at minimum using struct
there causes warnings.

Copied the license from the vnc plugin.

Compiles, but no testing yet.