[ENH] Move to pyproject.toml

[Moritz-Alexander-Kern]

This is a long one!

Original PR Body by Moritz

This PR moves Elephant from setup.py to pyproject.toml, aligning with PEP 518.

Key changes include:

• Introduction of pyproject.toml
• setup.py is still needed to compile the C-code for the fim module.

The reason is that we need OS dependend compile args which are dynamically defined in setup.py.
The pyproject.toml does not natively support conditional logic like os_name == 'windows' directly within the configuration. The TOML format is static and doesn't allow for dynamic evaluation of conditions based on the operating system.

To achieve platform-specific compile arguments, we have to use a Python script, such as a setup.py, which can programmatically determine the OS and apply the correct compile arguments.

In this context, as a configuration file for setuptools, setup.py is not deprecated, see also: https://packaging.python.org/en/latest/discussions/setup-py-deprecated/

---

Changes

pyproject.toml

• Created pyproject.toml with full project metadata (name, version, description, authors, license,
  URLs, classifiers)
• Moved to the SPDX licence expression format per PEP 639; removed the legacy License :: classifier string
• Switched to auto-discovery of packages via setuptools.find_packages(), this correctly picks up
  subpackages such as elephant.asset
• Added include-package-data = true so non-Python data files are included in the wheel install
• Added classifiers for Python 3.13 and 3.14
• Removed the long-unused six dependency
• Added viziphant to the tutorials optional-dependency group
• Dependencies are now declared via dynamic metadata, pointing back to the requirements/*.txt files. This keeps a single source of truth for dependencies

CI modernisation (.github/workflows/CI.yml)

• Merged pip test jobs: Three separate pip-test jobs (Ubuntu, macOS, Windows) merged into a single job with matrix.include, eliminating duplication
• Composite action: Added a composite action(.github/actions/restore-elephant-data) to avoid code duplication of restore action that restored cached elephant-dataset and is shared by the pip, MPI and conda jobs
• Enabled enableCrossOsArchive: true on the data cache so a cache built on one OS can be restored on
  another, this was added to build to both restore action and cache build action (.github/workflows/cache_elephant_data.yml)
• Bumped actions/cache/, actions/setup-python and actions/checkout to newer versions to fix issue with Node.js 20 actions being deprecated
• Made python version common accross jobs
• Per-job coveralls uploads + finish job: every job now uploads its own partial coverage report to coveralls, and a new finish job combines them at the end. This is to fix the earlier race condition where last coverall job was being usede to add PR comment.
• Added pytest-cov to requirements/requirements-tests.txt to avoid duplicate install steps for each job
• Fast CI mode: added a fast workflow-dispatch boolean input that limits the matrix to a subset of jobs
• Cleaned up workflow triggers
• Cleaned up the DOC BUILD and DOC TEST jobs

Environment files

• Refactored requirements/environment*.yml conda environment files to import from the corresponding requirements/*.txt files instead of duplicating the dependency list (system-level conda-only packages are still listed directly)

Bug fix for Windows path

• elephant/datasets.py: switched to local_file.as_posix() when constructing the ValueError message in download_datasets. On Windows, pathlib rendered backslashes, causing test_download_with_cache_dir_with_failed_integrity_check to fail because the assertion looks for the POSIX-style repo_path substring.

[coveralls]

coverage: 88.981% (+0.6%) from 88.424% — INM-6:enh/pyproject.toml into NeuralEnsemble:master

[CozySocksAlways]

Postponing addition of GPU tests with group runner available on jugit.fz-juelich.de to a different PR.

[CozySocksAlways]

Once PR approved, full CI should be triggered. (Run once for each PR approval)

[mdenker]

Thank you for this extensive PR, besides the comment about Python versions, it looks good to me so far -- I guess we need to see how the remaining CI jobs work out. I tested installs on Linux and Windows, which seem to work (including the fim module unter linux). However, the Python 3.8 issue that prevents installation of the package in a uv managed pip environment is something we should address.

[mdenker]

I feel like this line is giving me problems when doing the following:

• Create a new environment using uv venv
• Doing a uv pip install . in the elephant directory
• Calling uv run pythonm which gives me an error about dependencies not being satisfyable:uv responds that while I am currently using Python 3.14, the reuires-python indicates >=3.8. This is however incompatible with numpy 2. Therefore, uv fails, as it cannot resolve for any of the support python versions. I do not fully understand this logic, however, I guess we should drop 3.8 despite this being a hot fix. Or maybe I am missing something?

[CozySocksAlways]

Yup I could recreate this issue, it's an error from my side. The setup.py on our latest release has a requirement on python>=3.9, so I'll fix this now on the toml file.

As a side note, the latest release of numpy (2.4.6) has a requirement of python>=3.11, though lower version of numpy 2.0 still support python 3.9. So we should definitely fix this on our next minor release.

[CozySocksAlways]

Test duplication in case of pull_request_review trigger (i.e. both pull_request_review triggered tests pollutes the existing PR checks summary made with the checks triggered on PR synchronize) is the right behavior on GitHub and cannot be avoided as GitHub treats these two as separate events. Refer https://github.com/orgs/community/discussions/26940

As some form of relief, an if clause has been added on the setup job that makes it only run when pull_request_review is approved and not when a review comment is added.

[kohlerca]

The PR looks good to me. I tested locally in Ubuntu, using mamba and uv, and the installation succeeds.

[CozySocksAlways]

• Update names of tests for consistency (i.e with MPI and conda)