Based on behavior of Windows revealed by interactive tests added by Bruno Jesus.

Notice that I'm using floats instead of doubles in the new function,
mostly to be able to use struct wined3d_matrix and multiply_matrix().
At a rough estimate the precision should still be good enough.

While these won't be used by the shader, they potentially affect the
calculated offset for WINED3D_APPEND_ALIGNED_ELEMENT elements.

Instead of a fixed array of wined3d_shader_signature_element structures.
Shader model 4 shaders can have different semantics in a single
register, e.g.  v1.xy TEXCOORD0 and v1.zw TEXCOORD1, so having a single
wined3d_shader_signature_element structure per register isn't
necessarily sufficient.