opencv{3,4}: move cudatoolkit to propagatedBuildInputs

[tpwrules]

Motivation for this change

OpenCV's cmake files expect to be able to find the same CUDA toolkit they were built with when building packages that depend on OpenCV, even if they do not use CUDA. Fixes #144054 .

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandbox = true set in nix.conf? (See Nix manual)
• Tested via one or more NixOS test(s) if existing and applicable for the change (look inside nixos/tests)
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review wip"
• Tested execution of all binary files (usually in ./result/bin/)
• 21.11 Release Notes (or backporting 21.05 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
• Fits CONTRIBUTING.md.

[r-burns]

Technically, cudatoolkit should be a propagatedNativeBuildInput since it runs on the build platform. In any case, this is a clear improvement.

[hacker1024]

It seems that #221370 has re-introduced the issue that this solved. @ConnorBaker?

[ConnorBaker]

@hacker1024 I tried to reproduce the failure the original issue had but was unable to with

$ cat ~/.config/nixpkgs/config.nix 
{
  allowUnfree = true;
  cudaSupport = true;
  cudaCapabilities = [ "8.9" ];
  cudaForwardCompat = false;
}
$ nix build --impure -L github:nixos/nixpkgs/7fd504ce341d59b0ae2e8dc4469cc3ce3891eeb8#qimgv

(the commit 7fd504c is from earlier today).

We had some changes in how CMake-based projects handle/use CUDA redistributables which could be preventing this issue from occurring again: #233581.

Do you have a reproducer for the failure?