Properly leverage pre-commit cache

Currently the checks.yaml shared workflow is only run on PRs. That is problematic because the resulting cache is only shared within the PR. What we really want is a cache that is shared across PRs. Such a cache should generally be cheap to maintain, it just requires that we run pre-commit in our build.yaml so that we can have a cache tied to main that is available to all PRs and is keyed on the .pre-commit-config.yaml contents and is therefore updated only when we update that file in a PR. As one data point, currently the per-PR pre-commit caches are occupying >7Gb of the 10Gb per-repo GHA cache quota in cudf, so cleaning this up is is critical to unlocking usage of the cache for other purposes.

PRs to seed the pre-commit cache from main builds

is keyed on the .pre-commit-config.yaml contents and is therefore updated only when we update that file in a PR.

The official pre-commit GitHub action already does this: https://github.com/pre-commit/action

It's so small that it fits inline here, and we could probably inline something similar in shared-workflows if we wanted to avoid a dependency on it:

name: pre-commit
description: run pre-commit
inputs:
  extra_args:
    description: options to pass to pre-commit run
    required: false
    default: '--all-files'
runs:
  using: composite
  steps:
  - run: python -m pip install pre-commit
    shell: bash
  - run: python -m pip freeze --local
    shell: bash
  - uses: actions/cache@v4
    with:
      path: ~/.cache/pre-commit
      key: pre-commit-3|${{ env.pythonLocation }}|${{ hashFiles('.pre-commit-config.yaml') }}
  - run: pre-commit run --show-diff-on-failure --color=always ${{ inputs.extra_args }}
    shell: bash

ref: https://github.com/pre-commit/action/blob/1b06ec171f2f6faa71ed760c4042bd969e4f8b43/action.yml

For any repo where pre-commit run --all-files is self-contained (doesn't actually need to be in the conda env we set up by convention in check_style.sh), switching would be a quick win.

But that'd mean unwinding some of the conventions around how check_style.sh works and figuring out where to stick that action (in checks workflow, probably?, but idk about new job vs. new step).

I'm not exactly sure what you're proposing, but I implemented this "manually" in cudf in NVIDIA/cudf#24093 by just running the checks.yaml workflow in build.yaml. Does that track with what you'd expect?

Here's an example of a recent PR (NVIDIA/cudf#24101) where the first run of the checks.yaml pre-commit job hits the cache and doesn't need to re-setup the pre-commit environments: https://github.com/NVIDIA/cudf/actions/runs/34427112171/job/102714611203?pr=24101

For comparison, here are two random PRs along with their very first runs of the style check job:

🫠 I didn't actually know that checks.yaml was already saving and restoring the pre-commit cache, and apparently has been for 4 years: rapidsai/shared-workflows#20

ignore my comments, they all were made assuming that that wasn't already happening.

Got it never mind then! Yes, the issue isn't that we haven't been saving, it's that we've been caching in a way that doesn't share across PRs and also dramatically bloats our cache usage.

Seed pre-commit cache from main builds (#1086)

Seed pre-commit cache from main builds (#1100)

Seed pre-commit cache from main builds (#553)

Seed pre-commit cache from main builds (#314)

Seed pre-commit cache from main builds (#1244)

Seed pre-commit cache from main builds (#8729)

Seed pre-commit cache from main builds (#242)