Add dependency floor checks and generate conda recipe dependencies from pyproject.toml

[harryswift01]

Summary

This PR adds CI checks that the lowest allowed versions of our dependencies still install and pass the unit tests, and makes the conda recipe read its dependencies from pyproject.toml so the two can no longer drift apart.

Changes

Dependency floor checks:

• Add steps to the pull request unit job that resolve the lowest allowed dependency versions as wheels, reinstall them, and rerun the unit tests on Linux, macOS and Windows for Python 3.12–3.14.
• Add a weekly dependency-floors.yaml workflow, also runnable on demand, that runs the same check.
• Add a "Dependency Policy" section to the developer guide and a Dependency Floors badge to the README.

Conda recipe generated from pyproject.toml:

• Build the recipe's run: requirements from pyproject.toml, mapping art to python-art and lowercasing names.
• Take the waterEntropy pip pin from pyproject.toml too, instead of repeating it in the build script.
• This also fixes the stale numpy and rdkit ceilings in meta.yaml.

Raise floors pinned by waterEntropy:

• Raise the numpy, psutil, dask and distributed floors to 2.3.5, 7.1.3, 2025.7.0 and 2025.7.0.
• The lowest waterEntropy allowed (2.0.0) pins these versions exactly, so they are the oldest versions an install can actually have, and the lowest-version check now tests them.

Impact

• A dependency floor that no longer installs or passes the unit tests is caught on every pull request and in the weekly run.
• The conda package's dependencies always match pyproject.toml, so Renovate updates no longer leave meta.yaml behind.
• numpy older than 2.3.5, psutil older than 7.1.3, and dask or distributed older than 2025.7.0 are no longer supported.

[jimboid]

Looks like it will work to monitor the floor of our software as renovate only increases the ceiling. The conda recipe is a nice touch, but we should monitor the development of conda-pypi since that is likely going to solve some of these issues independently