Add new NVIDIA CUDA feature

[JP Ungaretti (jungaretti)]

Adds a new feature that installs Nvidia's CUDA and cuDNN libraries. These libraries enable tools such as TensorFlow to leverage an Nvidia GPU to perform highly-parallelizable tasks.

Motivation

This new feature would dramatically decrease friction for developers who use Codespaces to work on machine learning projects. Instead of copying, pasting, and executing this script, developers could simply add the new nvidia feature to their devcontainer.json and rebuild their codespace.

Changes

• Added a new Nvidia feature
• Added a test for the new feature
• Added two scenario tests for the new feature

[Chuck Lantz (Chuxel)]

There's a bit of a gap in the dev container spec for features that is needed here. Specifically, there's no way that I'm aware of for a feature to specify GPU usage - which I'd expect you'd want for this feature, correct? Is there any reason to use this feature without it?

EDIT: Raised devcontainers/spec#82

[JP Ungaretti (jungaretti)]

Specifically, there's no way that I'm aware of for a feature to specify GPU usage - which I'd expect you'd want for this feature, correct? Is there any reason to use this feature without it?

Not really. These libraries enable tools to leverage NVIDIA GPUs. That said, you can successfully install this feature without a GPU, but it won't be too useful.

[JP Ungaretti (jungaretti)]

This PR is ready for another set of eyes! Here are my notable changes:

• Added more helpful error messages when package installation fails. This feature prints an error message with the failed package (for example, libcudnn8=8.5.0.96-1+cuda11.7) and a link to NVIDIA's package repo.
• Added support for installing a certain version of CUDA and cuDNN. Unfortunately, NVIDIA's package names include their minor version number (for example, cuda-libraries-11-7). I've hardcoded the latest CUDA and cuDNN versions, but I'm open to more elegant solutions. This problem has existed since my first commit.
• Pivoted to always using NVIDIA's Ubuntu 20.04 package repo. I've found a lot of inconsistency among NVIDIA's Debian and Ubuntu repos. The Ubuntu 20.04 packages work perfectly on every image I've tested (Debian 10, Debian 11, Ubuntu 18, Ubuntu 20, Ubuntu 22). Another approach here would be to fail package installation with the more helpful error message.