I suspected that in this situation d3d writes to the z buffer, but the
test shows that I was mistaken

Add a test for this behaviour.

The render target should be left alone if the render target argument is
NULL. WineD3D calls should be done inside the critical section.

Turns out the original fix was correct for fixed function, but for the
wrong reason. The shader path was already correct. This fixes a
regresssion introduced by 932e95c1.

It seems that with Win2k3, the meaning of SO_REUSEADDR got closer to
BSD semantics. This is nice, as it allows us to finally work like some
version of Windows here.