Currently this worked by luck because it is leaking; the element holding
it never releases it.
Signed-off-by: Gabriel Ivăncescu <gabrielopcode@gmail.com>

The new test doesn't pass, but it doesn't crash either, and that's
enough to fix the WinZip installer.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=54071

lparam != 0 does not imply lparam_size is big enough for a MSG, so we
can end up manipulating memory past the end of the buffer.
Co-authored-by: Jacek Caban <jacek@codeweavers.com>

Update the desktop window size to match the current virtual screen rect.

Use a driver internal window message to dispatch updates to display devices from
the desktop window thread.

Store all incoming Wayland output state as pending and make it current
in a thread-safe way when the "done" event arrives. This enables other
threads, with proper locking, to safely read consistent and complete
Wayland output information.

Handle Wayland output events sent after process initialization,
and update the win32u display devices when handling these
events in the desktop window process.

Create a dedicated thread from which we read and dispatch Wayland events
(beyond the initial ones).

There is currently no benefit to dynamic allocation, and static
allocation allows us to simplify some aspects of initialization.

VPN interfaces often have ARPHRD_NONE in ifr_hwaddr.sa_data
but IFF_POINTOPOINT flag set.

It's less physically correct but closer to the native behavior.

We are not going to support multi-machine .so modules.

programs/explorer/desktop.c:104:16: warning: ‘hres’ may be used uninitialized in this function [-Wmaybe-uninitialized]
  104 |         return hres;
      |                ^~~~

Using TransformMatrixPoints is not needed and all values could
be taken from transformation matrix:
 - ShearX from m11, m12
 - ShearY from m21, m22
 - Translation mdx, mdy

The result could be calculated by taking destination points values:
 {{0.0, 0.0}, {1.0, 0.0}, {0.0, 1.0}}

and calculating it with GdipTransformMatrixPoints function:
  dst_to_src_points[0].X = dst_to_src.matrix[4];
  dst_to_src_points[0].Y = dst_to_src.matrix[5];
  dst_to_src_points[1].X = dst_to_src.matrix[0] + dst_to_src.matrix[4];
  dst_to_src_points[1].Y = dst_to_src.matrix[1] + dst_to_src.matrix[5];
  dst_to_src_points[2].X = dst_to_src.matrix[2] + dst_to_src.matrix[4];
  dst_to_src_points[2].Y = dst_to_src.matrix[3] + dst_to_src.matrix[5];

Improves efficiency by using addition instead of float numbers
multiplication for destination points.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=53947

With previous implementation, every iteration pointer value was
calculated by taking row and column of the image.
It needs multiply calculation. With current implementation,
pointer value calculation, is replaced with iterator, which takes
next pixel data.

It improves efficiency by using addition instead of multiplication
for iterating through points data.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=53947

Calculating transformation points and not using it it is a waste
of CPU time. With this patch the code was moved where it is used.
It is improving performance of GdipDrawImagePointsRect when
resampling is not needed.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=53947