Wheel generation. #3

Generate Python wheels with Arrow/Parquet dependencies

Description

Create distributable Python wheels for cpplink.

The Python bindings themselves are implemented, but generating a wheel is currently blocked by the C++ linking stage. cpplink depends on Apache Arrow and Parquet, and the wheel build needs to correctly resolve and package/link these native dependencies.

The main challenge is determining how Arrow and Parquet should be handled when building the wheel, particularly across different platforms.

Goals

• Build a Python wheel containing the cpplink Python bindings.
• Resolve the C++ linking against Arrow and Parquet correctly during the wheel build.
• Determine whether Arrow/Parquet libraries should be bundled with the wheel or provided as external dependencies.
• Ensure the resulting wheel can be installed into a clean Python environment.
• Test the installed package by importing cpplink and exercising Parquet functionality.

Areas to investigate

• CMake integration with Arrow and Parquet.
• Correct propagation of Arrow/Parquet include directories and libraries to the Python extension.
• Runtime library discovery (rpath, shared libraries, etc.).
• Wheel packaging of native libraries if required.
• Platform-specific behaviour on macOS and Windows.
• Interaction with pybind11 and the Python extension module.
• Whether Arrow/Parquet should instead be provided through Conda dependencies.

Acceptance criteria

• A wheel can be generated successfully from a clean build.
• The wheel installs successfully into a clean Python environment.
• import cpplink works without requiring a local development build.
• Arrow/Parquet symbols are correctly resolved at build and runtime.
• Parquet read/write functionality works from Python.
• The dependency strategy for Arrow/Parquet is documented.
• The approach can be extended to Windows wheels.