Mono calls GdipMeasureString with a height so large that rounding and
converting it to an integer gives a negative result.

This fixes a regression introduced by 10f58c14.

This patch should fix the regression(s) caused by my recent button painting
patches. The problem was that DrawFrameControl() was modifying the rectangle,
and one of my patches stopped calling DrawFrameControl() in case of ODA_FOCUS
action, so drawing an entire button and only focus started to differ, which
was hidden by the custom clipping (that was removed by my last patch).

MSDN says the contrary.

wined3d: Separate AMD GPUs in D3D9 and D3D10 models, this allows for a more correct fall back when the GPU table isn't up to date.

D3D10_CAPABLE should work on any AMD OpenGL 3.0 driver (those came out in september 2008) and on OSX 10.5 and up.

wined3d: Separate AMD Radeon HD2300 from other HD2xxx/HD3xxx GPUs since it is D3D9 instead of D3D10.

Apparently matrices in gdiplus don't work the way I think they do.

This will make it possible to use the same codepath for non-bitmap graphics
objects and to create a software implementation for other drawing operations.