📦 Split heavy deps into optional extras and slim the wheel

[dariocazzani]

What does this PR do?

Ships the correct minimal dependency set for pip install devol. Before this PR, installing the library pulled in torch, torchvision, gymnasium, and matplotlib — hundreds of MB — even though the runtime code in src/devol/ never imports any of them. After this PR the base install is just numpy, pydantic, and pydantic-yaml; the heavy stack moves under the examples extra where the cartpole/MNIST/two-peaks demos actually use it.

Details

Required dependencies reduced to 3 packages

• numpy>=1.26, pydantic>=2.0, pydantic-yaml>=1.6
• Dropped unused pydantic-settings (leftover import from an abandoned cleanup branch, never referenced in code)

Extras restructured

• examples — matplotlib, gymnasium[classic-control], torch, torchvision
• benchmark — matplotlib, rich, tqdm
• dev — ruff, mypy, pytest, twine (added twine so the build CI job can validate metadata without an ad-hoc install)
• all — meta-extra that resolves to devol[dev,examples,benchmark]

Wheel narrowed to src/devol/ only

• examples/ was previously shipped inside the wheel even though it's demo code; now it stays in the repo but out of the package
• Verified: fresh devol-0.1.0-py3-none-any.whl contains 9 files under devol/ + metadata — no examples, no benchmark, no tests

Pytest pythonpath config

• tests/test_ci_n_peaks.py imports from tests.ci.n_peaks. With examples/ previously shipped as a top-level wheel package, the editable install accidentally exposed the repo root on sys.path, which is where pytest found the tests namespace. After narrowing the wheel, that side effect is gone, so pytest gets an explicit pythonpath = ["."] and testpaths = ["tests"] config.

README and CHANGELOG

• Added an Installation section to the README explaining the extras
• Changelog entries cover the dep split, the wheel narrowing, and the earlier CI/typing work from PR 🏗️ Split CI into lint/typecheck/test/build and type-clean src/devol #10

Verification

Minimal install (no extras) runs the algorithm end-to-end:

uv sync --no-dev --no-extra dev --no-extra examples --no-extra benchmark --no-extra all
python -c "from devol import DiffusionEvolution, DiffusionConfig; import numpy as np; \
  algo = DiffusionEvolution(DiffusionConfig(population_size=32, num_steps=10, param_dim=4, sigma_m=0.5, seed=42), \
                             lambda x: float(-np.sum(x**2))); \
  algo.run(); print('best:', algo.get_best_individual()[1])"
# → best: -0.1714

python -c "import torch"
# → ModuleNotFoundError: No module named 'torch'   ✓

All CI gates green locally:

• uv run ruff check ✓
• uv run ruff format --check ✓
• uv run mypy src/devol ✓
• uv run pytest ✓ (15 passed)
• uv build + twine check dist/* ✓