spec/glsl-1.20: Use min and max with NaN
In IEEE-754 and OpenCL, min and max with a single NaN parameter are required to return the other paramter. GLSL does not have this requirement, but numerous applications depend on this behavior. These tests try to exercise this behavior in a way that should be compatible GLSL implementations that cannot produce NaN or Inf.
Reproduces the failure in mesa#4254 (closed).
I have only tested this on Intel GPUs. It's possible that these have problems on R300/R400 generations of hardware... I'm also not sure if those drivers expose GLSL or only the older assmebly shading languages.