CUDA: Update runtime and SDK to v12.8.

[maleadt]

No description provided.

[stemann]

Is this the time to add the lib/lib64 symlink: ln -s $prefix/cuda/lib $prefix/cuda/lib64?

[giordano]

Why? Note that symlinks cause full duplicates on Windows

[maleadt]

I only added it to the Linux build for that reason, so @stemann maybe you could elaborate why you need that.

[stemann]

Cf. the discussion we had on slack in #binarybuilder - it was needed for the Faiss recipe for CMake to find a functional nvcc - I copied it from the cufinufft recipe.

[maleadt]

Only having the symlink for Linux seems strictly worse though, because packages will now still need a CMake workaround for some of the platforms.

Alternatively, @giordano, could we still keep the symlink for Windows since the SDK JLLs are typically only installed on Yggdrasil? FWIW, regular CUDA toolkit installations seems to come with it:

❯ ls /opt/cuda -la | grep lib
lrwxrwxrwx 1 root root    5 Nov 21 19:51 lib -> lib64/
lrwxrwxrwx 1 root root   24 Nov 21 19:51 lib64 -> targets/x86_64-linux/lib/

[stemann]

It might also not be the time - if it means that CUDA 12.8+ must be handled differently from CUDA 11.4-12.7. It could also be postponed to sometime when a rebuild of all the SDK versions are needed - though I’m not sure if there is any chance of that happening… (could BB2 trigger a change?)

[maleadt]

OK, let's revert it then for now.