Add some official Docker images

[francoishernandez]

Several objectives, including:

• facilitate setup and onboarding for any user, potentially simplify some docs/tutorials
• facilitate reproducibility, issue reporting and bug fixes

This is a first version, we might want to adapt some things down the line, e.g.:

• build for specific / multiple torch version(s) (currently defaulting to latest available)
• faster/lighter build (relying on some official nvidia images? multi-stage? caching?)
• some inference specific image(s), which would need less dependencies and hence could be significantly lighter;
• test and adapt for other platforms (tested on a fairly standard linux amd64 server only for now)

Build is automatically triggered for every push on this docker branch to faciliate testing the workflow, but afterwards it will be handled via workflow_dispatch. It is intentionnaly decoupled from the release workflow, to facilitate manual builds down the line (e.g. for specicific torch/cuda versions, etc.)

[vince62s]

maybe add instructions on Docker usage in README ?

[francoishernandez]

Latest commit 117be02 adds a bit of doc in the readme, as well as the install of all the optional requirements in the Dockerfile. This first version of the image(s) will serve as a kind of "exhaustive" dev platform, and we might create some lighter alternatives down the line.
Will merge and trigger the first build on v3.4.3.

[l-k-11235]

Merci !

[anthdr]

Pytorch images provided by nvidia already have everything built-in (including apex) which simplify even more the creation of the image, but seem to offer less flexibility regarding cuda & pytorch version (and I'm not sure the whole image can pass in the github action servers).
https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch
https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html#framework-matrix-2022