Whenever a new GPU generation comes out, Blender always needed an update to ship new pre-compiled binary kernels for that generation. That has some drawbacks, including e.g. that older Blender versions won't work on those GPUs (unless adaptive compilation is used, but that's really for advanced users).

This patch changes the discovery of pre-compiled kernels, to look for any PTX kernels, even if they don't match the current architecture version exactly. This works because the driver can JIT-compile PTX generated for any architecture less or equal to the current one. For this to be effective the buildbot would ideally not only build CUBINs for the various architectures like it does now, but also include PTX builds as a fallback (at least one, e.g. "compute_75" to e.g. allow Ampere to work, without the explicit need to introduce the CUDA 11 toolkit into the build pipeline yet). Haven't included that change yet, since I wanted to discuss first.

I think this brings the best of both worlds together: CUBIN for fast load times on selected architectures and PTX for compatibility on all others.

Also bumped up the target architecture for the OptiX kernel as discussed in D8063.