Long IRs are more expensive than short IRs, for sure.
So 7 x 256 means that it’s buffering 1792 samples, but the callback rate is higher, so the audio render feels more responsive because it’s getting updated data from the audio thread at a higher frequency.
This helps things like panning feel more responsive at higher buffer sizes.
Convolution, even though it’s been around for a while, is very much a next gen feature because it’s always been expensive. Both of these features should probably be treated as next-gen features. So from a performance stand-point, I would suggest thinking about your min spec. Maybe that means using an algorithmic reverb instead of convolution reverb on lower tiered machines or platforms.
However, as we look toward the future, the ability of the average machine to handle things like convolution and ambisonics is only going to get better and better. 