gdi32: Japanese Windows had a default fallback Font SystemLink of Microsoft Sans Serif. Implement that for Japanese.

The fog test used start = 0.9 and end = 0.1. This is perfectly ok in
theory, but some Windows drivers do not like it. Since this test tests
how fog is calculated with Vertex and Pixel shaders put the fog range in
the normal order(start = 0.1, end = 0.9), this fixes a number of fog
failures on ATI cards on Windows.

The drivers for some radeon cards are not prepared for the things the
fixed function type test does and fail inconsistently. Thus skip the
color comparison if one of the two draws fails, and do not write an
extra complaint about the inconsistency

The state manager calls GetMemory before it calls PreLoad on the VBO,
and PreLoad depends on finding the VBO in the strided vertex structure.
This can cause problems on the first draw which creates the vertex
buffer, because the first PreLoad creates the vbo and attempts to
convert with an incorrect strided structure.

vertex_desc.Pool should obviously be used in the 
WINED3DRTYPE_VERTEXBUFFER case instead of index_desc.Pool.

The base version won't suffice anymore as it is not able to upload
palette changes to the drawable in an efficient way for both GDI and
GL. Further the LoadLocation code in RealizePalette isn't needed for
the GDI version as in all cases it works on system memory.