enable pre-commit

[antoinejeannot]

This PR (re)enables pre-commit!

It has been discussed in a few issues & PRs now (#120, #119, #431), with good reason since it definitely improves the developer's workflow.

I've already ran into a few linting & formating issues myself with just a few PRs 😅

A few points:

• to share the same behavior locally and in the CI (which leads to pain), ruff was pinned in both the pyproject.yaml and .pre-commit-config.toml, let's bump it from time to time or setup a dependabot
• to assess quality, only pre-commit is used (more robust thanks to its dependencies isolation), but is iso-functional with a plain ruff xyz since both rely on the pyproject.toml for configuration
• ruff is still mentioned in pyproject.toml so that the developer and its editor code editor can pick it up from the virtualenv and apply linting / formating
• make style is redundant with make quality, so I deprecated it with a log for those who use it
• the quality workflow, thanks to some caching, can save up to 15s per run (when cache hit), this might definitely help the tests workflow as well
• ruff is now configured with the lower bound of supported versions (3.10, instead of 3.12) as recommended, which probably led to some flaky linting & formating issues
• the check tests in ci is now a .pre-commit hook

I have another PR which adds a few needed additional rules to the linter but will lead to major merge conflicts across all opened PRs lol I'll open it in a second time.

Thanks for reviewing @aymeric-roucher @albertvillanova

[antoinejeannot]

pre-commit manages its own isolated venv, hence no more env management.
Also no need to install smolagents for lint & formating, only pre-commit and ruff are needed

[antoinejeannot]

this action uses some caching under the hood, which speeds up ci actions

[antoinejeannot]

Now a pre-commit hook

[albertvillanova]

I would prefer to stay aligned with all the other huggingface open source libraries.

Additionally, previous attempt to use pre-commit had a compatibility issue with the CI quality job (handled with ruff), if I remember correctly.

[antoinejeannot]

Several other HF repos also seem to leverage pre-commit (datasets, accelerate, peft, tgi etc.).

I feel your concern about compatibility issues, and that's actually what drives this PR:
pre-commit actually runs its hooks inside their own venv, bringing robustness and reproducibility to the table.

There are a few reasons pre-commit might have been flaky so far in smolagents:

• ruff is not stable and must be pinned, otherwise ruff 0.9.0 will differ from ruff 0.9.1 (patch): this explains why there might be some issues between the CI and local environments right now
• ruff must be pinned in pre-commit-config.yaml to the same version as in pyproject.toml, otherwise there could be a difference as well
• ruff's version in .pre-commit-config.yaml is outdated
• pre-commit base hooks must run before ruff, otherwise they could introduce changes after ruff (depending on the hook) that can be committed

I am confident this will make the CI/CD more reliable, as well as smolagents more maintainable. WDYT ?