sped up nexuiz rtlights benchmark by 4% by optimizing box cull on models to skip...

sped up nexuiz rtlights benchmark by 4% by optimizing box cull on models to skip triangle box cull when surface box is entirely inside light box
restructured R_Shadow_ConstructShadowVolume loops a bit to make for less stalls, this seems to have only netted a 0.5% speed gain or so

git-svn-id: svn://svn.icculus.org/twilight/trunk/darkplaces@4566 d7cf8633-e32d-0410-b094-e92efae38249