fix: better CUDA detection

[delehef]

What ❔

This PR improves the detection of the CUDA toolkit under Linux, and the detection of its version under all OSes.

Why ❔

Under Linux, the CUDA detection is hardcoded to look into /usr/local/cuda, which is not guaranteed to be the default installation folder. The new method first looks for the CUDA_PATH env. var., then default to /usr/local/cuda if it is empty.

The current way of finding out the installed CUDA version is to look for a version.json file in the CUDA installation path. This file is not guaranteed to exist. However, nvcc is guaranteed to be installed with the toolkit; so the new method is to parse its output to find the exact version there.

Checklist

• PR title corresponds to the body of PR (we generate changelog entries from PRs).
• Tests for the changes have been added / updated.
• Documentation comments have been added / updated.
• Code has been formatted via zk fmt and zk lint.

[robik75]

AFAIK unfortunately there is no canonical way to determine where CUDA toolkit is installed and what the version of that installation is.
Parts of https://cmake.org/cmake/help/latest/module/FindCUDAToolkit.html can be taken as inspiration.
Every somewhat recent CUDA toolkit installation now includes the version.json file and all the versions that we care about (11.x and up) included it for sure. We do not support older CUDA versions anyway.
Did you encounter a scenario where using version.json did not work for you?

[delehef]

AFAIK unfortunately there is no canonical way to determine where CUDA toolkit is installed and what the version of that installation is.

That's true, and that's most unfortunate :/

Did you encounter a scenario where using version.json did not work for you?

AFAICT, the version.json file in only guaranteed to exist in manually installed setups. I can say that e.g. for Ubuntu, Debian, and NixOS, it does not appear when installed from the canonical package manager.

[robik75]

Hmm, my experience it that installing CUDA manually on Ubuntu following the official CUDA installation procedure (that uses apt with nvidia's package repository) results in version.json being placed in the root of the CUDA installation.
Either way, we can proceed with your proposed changes, as long as the CI works, it should be fine.

[delehef]

version.json being placed in the root of the CUDA installation.

Absolutely, but only for the manually installed version. If ones uses apt to install it, then version.json is not present.

Thank you for the feedback, I addressed your comments.

[robik75]

The PR branch seems to be out of date with the base branch. Can you please rebase?

[delehef]

Done