Keep the dependency specs consistent with the conda recipe

[xylar]

mpas_tools describes its dependencies in four places — recipe/recipe.yaml, pyproject.toml, dev-spec.txt and pixi.toml — and they had quietly drifted apart. This PR fixes the discrepancies and adds a test so they cannot drift again.

Fixes

recipe/recipe.yaml is treated as the source of truth for the run requirements:

• pyproject.toml was missing the >=2.0,<3.0 constraint on numpy and the !=0.20.0 constraint on scikit-image
• pyproject.toml was missing h5py entirely, even though it is imported by mpas_tools.viz.mpas_to_xdmf.io and is available on PyPI
• pyproject.toml, dev-spec.txt and pixi.toml were all missing the >=0.10.0 constraint on (python-)igraph
• dev-spec.txt still asked for geometric_features >=1.0.1 rather than >=1.6.3
• dev-spec.txt and pixi.toml still asked for python >=3.10 while requires-python in pyproject.toml is >=3.11
• recipe/recipe.yaml now skips building for python <3.11, matching requires-python and the conda-forge feedstock

New test

conda_package/tests/test_dependencies.py checks that every run requirement of the conda recipe appears with the same version constraints in the other three files. dev-spec.txt and pixi.toml may still list extra packages for development, testing and building the documentation, but pyproject.toml is also checked in the reverse direction so that it cannot list a package the recipe is missing.

A few details worth noting:

• Package names are normalized following PEP 503, with a small table for the conda-forge names that differ from their PyPI equivalents (matplotlib-base → matplotlib, python-igraph → igraph).
• geometric_features and nco are not on PyPI, so they are skipped for pyproject.toml only.
• Version constraints are compared as unordered sets of clauses, so ordering and spacing do not matter.
• Python is checked against requires-python from pyproject.toml rather than the recipe, since the recipe deliberately leaves python unconstrained and lets conda pin it to the version being built. A constraint in the recipe is still checked if one is ever added.
• The tests skip when the spec files are absent, which is the case when the test suite runs during a conda build.

The test imports yaml, so pyyaml was added as a development dependency rather than relying on it arriving transitively via pre-commit.