cudaPackages: consider splitting libnvcuvid out from nvidia_driver

[SomeoneSerge]

• Verify if nvcuvid is actually hardware-depended and needs to be deployed impurely, like libcuda.so
• If it's not, split it out into a separate output or derivation
• Make other libraries link to it directly through Runpath

Motivated by #218859

[Kiskae]

❯ readelf -p .rodata /run/opengl-driver/lib/libnvcuvid.so | grep version
[5c97e0] Library version doesn't match driver version [NV_ERR_LIB_RM_VERSION_MISMATCH]

Pretty much all the nvidia userspace libraries are version-locked to the associated driver version, so this won't work.

[SomeoneSerge]

Ah, so that's how this works! I guess that settles it

[nixos-discourse]

This issue has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/python-dll-import/22594/13