Fix pinning wrong numpy version

[kounelisagis]

Right now, there is no single source of truth for dependencies. Until this is done, let's fix the required versions of numpy found in multiple project files. There are basically 5 different places where numpy version requirements should be specified:

1. requirements.txt
2. requirements_dev.txt
3. requirements_wheel.txt
4. pyproject.toml -> build-system -> requires
5. pyproject.toml -> project -> dependencies

They can be merged into 3 categories:

• build dependencies: pyproject.toml -> build-system -> requires and requirements_wheel.txt are the requirements when building wheels (https://setuptools.pypa.io/en/latest/userguide/dependency_management.html#build-system-requirement). For python 3.8 wheels have to be built against numpy 1.17 because we want support for >=1.17. NumPy's ABI is forward but not backward compatible before 1.25. This means; binaries compiled against a given version of NumPy will still run correctly with newer NumPy versions, but not with older versions (https://numpy.org/devdocs/dev/depending_on_numpy.html). For aarch64 and arm64 architectures the first version of numpy with available wheels is 1.19 and 1.21, and thus we set this. For python 3.9+ wheels can be built against latest numpy (like other big projects are doing: https://github.com/pandas-dev/pandas/blob/58461fef2315d228d08f65f8a9430e9294d65b31/pyproject.toml#L11)
• runtime dependencies: pyproject.toml -> project dependencies and requirements.txt are the requirements used to run the package (https://setuptools.pypa.io/en/latest/userguide/dependency_management.html#declaring-required-dependency). For python 3.8 whatever greater than numpy 1.17 (build version) is fine (for aarch64 this is 1.19 and for arm64 this is 1.21) and for python 3.9+ we want something greater than numpy 1.25 since NumPy C-API is exposed in a backwards compatible way by default.
• development dependencies: requirements_dev.txt. This shouldn't be much of concern since it's what we are installing to build and run locally. If we use the runtime dependencies it should be safe.

---

Testing.

Python 3.8 (macos arm64)
Build: numpy-1.21.6
Installed with wheel: numpy-1.24.4

Python 3.8 (ubuntu)
Build: numpy-1.17.5
Installed with wheel: numpy-1.24.4

Python 3.9 (macos arm64)
Build: numpy-2.0.0rc2
Installed with wheel: numpy-1.26.4

*Wheels were installed in a fresh conda environment using pip install --no-cache-dir.

---

[sc-48530]