Fix skala.ase import failure on CPU-only environments

[gncs]

Root cause

The docs build has been failing since #58 because docs/ase.ipynb cannot execute the very first cell:

from skala.ase import Skala

The CI traceback (myst_nb.core.execute.base.ExecutionError) hides the underlying notebook error; the linkcheck step is just the last to surface it (the HTML build hits the same failure a few seconds earlier).

src/skala/ase/calculator.py does:

try:
    import skala.gpu4pyscf as skala_gpu
except ImportError as e:
    skala_gpu = None
    gpu4pyscf_import_error = e

But src/skala/gpu4pyscf/__init__.py had, at module level:

if not torch.cuda.is_available() and find_spec("pytest") is not None:
    import pytest
    pytest.skip(
        "Skipping gpu4pyscf doctests, because CUDA is not available.",
        allow_module_level=True,
    )

pytest.skip(...) raises pytest.outcomes.Skipped, which derives from BaseException, not ImportError. So on any host where (a) CUDA is unavailable and (b) pytest is installed — which is every CI runner and every dev env that installs .[dev] or .[doc] — the except ImportError does not catch it, and import skala.ase fails with Skipped. This breaks both the docs notebook and any CPU-only use of the ASE calculator.

Repro on main:

$ python -c "from skala.ase import Skala"
...
File "src/skala/gpu4pyscf/__init__.py", line 19, in <module>
    pytest.skip(...)
Skipped: Skipping gpu4pyscf doctests, because CUDA is not available.

Fix

Two coordinated changes.

1. src/skala/gpu4pyscf/__init__.py — replace the module-level pytest.skip with a plain ImportError. The skala.gpu4pyscf package genuinely requires CUDA, so "module is unusable" is best expressed as ImportError, which is the standard signal callers already handle (including calculator.py's try/except ImportError). The find_spec("pytest") gate is dropped: it was both insufficient (any env with pytest installed triggered the bug) and unnecessary (the module being unimportable on non-CUDA hosts is correct independent of pytest).

2. src/skala/conftest.py (new) — the original pytest.skip was, despite the broken mechanism, trying to do something legitimate: prevent pytest --doctest-modules --pyargs skala (used by .github/workflows/test.yml) from failing collection of the gpu4pyscf modules on CPU runners. After change (1), pytest collection sees the ImportError and errors out. Restore the intended behavior with a collect_ignore_glob that skips gpu4pyscf/*.py when CUDA is unavailable. The conftest is placed at src/skala/ rather than inside src/skala/gpu4pyscf/ because, as a regular package, the inner directory's conftest would itself be imported as skala.gpu4pyscf.conftest, triggering the failing __init__.py before pytest could read it.

Dedicated tests in tests/test_gpu4pyscf_*.py already use their own pytest.skip(allow_module_level=True) guard, which is the correct place for that pattern (inside test modules, where pytest does the importing).

Verification

• from skala.ase import Skala works on a CPU-only env.
• The ASE notebook's GPU cells (Skala(..., device="cuda") and calc.set(device="cpu")) work, because calculator.py catches the ImportError and only raises at calculate() time if device="cuda" is actually invoked.
• sphinx-build -b html docs docs/_build/html succeeds end-to-end.
• pytest --doctest-modules --pyargs skala tests/ → 159 passed, 4 skipped on CPU-only env.
• ruff format, ruff check, mypy pass on both changed files.