We should be using only non-perceptual roughness in shader code.

It makes the code less confusing, and hopefully saves registers.

Iirc I think the only things using it are the F_AB split-sum approximation that we copied from Unreal, and environment_map.wgsl (which might make sense to keep as perceptual roughness).

EDIT: I'm not confident that F_AB is even meant to be using linear roughness.