🐳 Docker update + Simplify Jobs doc

[qgallouedec]

Our new simple docker image mostly intended to be used with trl jobs

Note that the image is already pushed to https://hub.docker.com/r/huggingface/trl-source-gpu (soon rename to just "huggingface/trl")

[qgallouedec]

in theory, we need would need to build

• dev:
  • every commit on main
  • every new release of a dependency
• stable:
  • every patch release
  • every new release of a dependency

however, since we typically make a few commits per day on main, let's simplify things and just build both images when a new commit is made on main.

[sergiopaniego]

Regarding the Training using Jobs guide, I can see that we'd need to update it adding:

• Details about the new trl Docker image with dependencies
• Details about TRL Jobs wrapper

Since it's just a small update, I can include it in this PR shortly. If you’d rather merge sooner, feel free, and I’ll open a new PR 😄

[qgallouedec]

I took the opportunity to simplify the jobs documentation:

• if it was not specific to training/trl, I redirected to HF Jobs doc (env var, hardware, timeout etc)
• add some warnings about important things like pushing the model and set a proper timeout
• remove the section on trackio as we now have a dedicated section in the doc for it
• made the training script part a bit more general (script don't necessarily come from the repo
• moved up trl-jobs, as it's the simpler solution so I expect user to go further in the doc for further customization (you don't want the user to discover at the end that a simpler solution existed)
• move the docker part down as it's probably more for power users to choose their own docker image

I'll merge this PR, as it's been opened for a while now, and I want the docker image to be operating. we can still refine in the future if needed

[qgallouedec]

for scripts, we don't want relative imports