Evaluate splitting bounces into dedicated diffuse and specular terms

[w23]

Alternatively to trying to estimate and mix both.
This can help with:

1. thread coherency: all threads are doing the same work, as opposed to half being diffuse and another random half being specular.
2. more bounces: it will be natural to know what channel we're writing even if we're N bounces deep

Note that fully diffuse and fully specular models are still fine -- just no compute will happen for them, which is cheap.