[API Compatibility] migrate workflows to github actions

[Manfredss]

PR Docs

[CI] Add GitHub Actions CI suite + PR /re-run action

Summary

Introduces the full GitHub Actions CI for PaConvert under .github/workflows/
plus a reusable composite action under .github/actions/. This covers code
style, API consistency, install/packaging, and CPU / GPU / distributed unit
tests, and adds a comment-driven workflow re-run mechanism for PRs.

Workflows added (.github/workflows/)

File	Workflow / Job	Runner	Purpose
codestyle.yml	CodeStyle Check / pre-commit	CPU* (ubuntu-24.04)	Runs pre-commit run --all-files (pre-commit 2.17.0)
consistency.yml	Consistency Check / Common API Consistency Check	CPU* (ubuntu-24.04)	Runs tools/consistency/consistency_check.py; uploads log on failure
Install_check.yml	Install Check	CPU* (ubuntu-24.04)	Builds the wheel (sdist bdist_wheel), installs it, runs paconvert -V / --run_check
Unittest_check.yml	CI Unittest	CPU* (ubuntu-24.04)	CPU pytest suite with reruns + --lf retry; uploads pytest.log on failure
Unittest_check_gpu.yml	GPU Unittest Check	self-hosted GPU (group: PaConvert)	GPU pytest suite (CUDA 11.8 torch/paddle)
Unittest_check_distribute.yml	Distributed Unittest Check	self-hosted GPU (group: PaConvert)	2-GPU distributed tests via run_and_compare.py
(CUDA_VISIBLE_DEVICES=0,1, --nproc_per_node=2)
Modeltest_check.yml	Modeltest Check	self-hosted GPU (group: PaConvert)	Converts model code-set from /workspace/torch_project, runs
tools/modeltest/modeltest_check.py
rerun.yml	Rerun Workflows	ubuntu-latest	Re-runs CI jobs from a PR comment (see below)

* CPU jobs currently run on ubuntu-24.04 with a # TODO: replace with the actual self-hosted CPU runner group name placeholder — see Follow-ups.

Composite action added (.github/actions/rerun-workflow/)

• action.yml — composite action Rerun workflow; inputs GITHUB_TOKEN,
  OWNER, REPO, PR_ID, JOB_NAME.
• rerun.sh — resolves the PR head SHA via the GitHub REST API and either
  re-runs all failed jobs (rerun-failed-jobs) or a specific job by name.

Comment-driven re-run

On a PR, the PR author can comment to re-trigger checks:

/re-run all-failed · codestyle · consistency · install-check
· cpu-unittest · gpu-unittest · distributed-unittest · model-test

(Guarded by github.event.comment.user.login == github.event.issue.user.login.)

Design notes

• Triggers: pull_request + push to master, develop, release/**,
  with paths-ignore for **/*.md, docs/**, LICENSE, .gitignore,
  .pre-commit-config.yaml.
• Concurrency: per-workflow/per-PR group with cancel-in-progress: true.
• Least privilege: permissions: contents: read everywhere; rerun.yml
  additionally gets actions: write.
• Reliability: Paddle installed from nightly with --timeout/--retries;
  GPU torch/torchvision pulled as pinned cu118 wheels from
  paddle-paconvert.bj.bcebos.com; pytest retried (--reruns=3, then --lf);
  failure logs/artifacts uploaded with 14-day retention.
• Nightly cron is scaffolded but commented out in every workflow.

Self-hosted runner prerequisites

• Runner group PaConvert online (it currently contains the paddle-gpu
  runner).
• Distributed job requires a host with ≥2 GPUs.
• Modeltest job requires the model code-set prepared at
  /workspace/torch_project on the runner.

Follow-ups

• Point the 4 CPU jobs (CodeStyle, Consistency, Install, CI Unittest) at the
  self-hosted CPU runner group (replace the ubuntu-24.04 placeholders).
• Enable the nightly schedule cron once the runners are validated.

Test plan

• CPU jobs (codestyle / consistency / install / CI unittest) pass on a PR.
• gpu-unittest, distributed-unittest, modeltest are picked up by the
  PaConvert group and complete.
• /re-run <job> comment re-triggers the targeted job(s).

PR APIs

N/A