Fixes:
1. Mixed use of `sample_reflection` and `sample_refraction`.
`sample_reflection` is intended for when only reflection is required.
For refraction, `sample_vndf` or `sample_refraction` should be used.
2. Wrong weight when accumulating the contribution. Previously
`brdf * NV / (fresnel * pdf)` always evaluates to `GL`, but with the new
technique of bounded VNDF sampling this is not true anymore. Fixed by
adding a field `weight` in the struct `BsdfEval`.
3. Schlick's approximation of the fresnel factor is
`F + (1 - F) * (1 - cos(theta))^5`, but BSDF LUT was using
`cos^2(theta)`, which was incorrect.
313a7d6236