Implement ray differentials?

[snogglethorpe]

Are these worth the effort?

Everybody seems to do them, and the extra info available can help choose the right amount of e.g. texture sampling to do.

However they mainly seem to be useful when using very low sample rates at the top-level, and we often want to use high sample rates for other reasons, e.g. to make depth-of-field blur look reasonable. Maybe it's better to just focus on making high sample rates faster?