What changed in EDuke32, build by build

Allow all drivers for now; if anything, we should start maintaining a blacklist now that most Intel chips have decent enough support to be able to run stuff satisfactorily.

This was broken ever since updatesprite() had been forked off drawsprite() and would cause static lights to never hit static sprites until one or the other changed, after which the light would always get culled against outdated sprite geometry.

Since updatesprite() only happens once for static sprites, avoiding light culling if we hit it in a shadow pass means that sprite will never be lit as long as that happens.

With VBOs enabled, the virtual pointer to the geometry of frame 0 of a model is always NULL. This was confusing the polymer code responsible for deciding whether to enable smooth animations or not, which meant that any animation that would loop to frame 0 would do it as if animation interpolation was disabled, which looks very choppy.

We pack our own copy of glext.h to avoid build breakages on older environments, and having gl.h include its own was defeating all that.

This also fixes a warning when building for Windows.

Getting the texture handles without requiring invalidation means we can handle ART-animating tilenums without impacting performance too much.

Polymer has a hack that snaps horiz values of spotlights between 90 and 110 to 90 and 110 to avoid triggering aliasing artifacts with perfectly horizontal spotlights. Propagate that change all the way to the spotlight SE sprite so that the code won't think that the horiz is always different from the light and mark it invalid each frame.

This won't actually do anything until the a debug bit is added when creating the context, but that requires support for GLX_create_context which SDL doesn't support. I'll add support for WGL_create_context to winlayer in a bit as a stopgap. Also updates our local copy of glext.h with a fresh one from the Khronos registry.

We need to Y-invert the projection texcoords when sampling from the lightmap.

Sprites would stick with their old material on global palette change. This goes through all prsprites and clobbers their CRCs so that an alternate skin / tint / highpal can be applied accordingly.

The winding order of faces is different if only flipping about X or Y, and we were unconditionally culling CW faces.

Keep track of which side is getting culled, swap around mirrors, swap around model drawing.

In cases where performing occlusion queries for multiple red walls leading to the same sector, we would sometimes leak a number of query objects. This was apparently harmless on NVIDIA drivers, but would leak tons of memory on ATI.

The code was using the Polymer-cached version of the sector cstat to determine the translucency of its diffuse material, meaning it would never get it right on the first try; that's what you get for having completely different variables named 's' and 'sec'.

We normally enable backface culling always, but turn it off for two-sided sprites; we need to make sure it's re-enabled after that as sector masks are drawn after.