user32: When building linedefs prev can be less than 0 if the format width is less than the width of a single character.

This avoids either loading sysmem surfaces into video memory (wasting GPU
memory) and then doing a blit on the GPU, or downloading GPU surfaces to
sysmem, doing a blit on the CPU, and likely uploading it again. This mostly
matters for ddraw and d3d8 applications, d3d9 already has to go through
wined3d_device_update_surface() to transfer data from sysmem surfaces to GPU
surfaces.