Enable Axom's Python to be installed via uv/pip

[kennyweiss]

Summary

• This feature PR packages Axom's Python bindings as a wheel that can be installed via uv or pip
  • Our current Python interface is run through a script (run_python_with_axom.sh) that adds the necessary dependencies to PYTHONPATH
  • This PR adds another option to generate a scikit-build-core project over an installed Axom. The generated Python wheel can be installed into a uv or pip virtual environment
  • It also adds a CI job to build and test the wheel via uv from an already built Axom installation (the gcc docker CI)
• This PR also improves the Sidre bindings to (a) fix type helpers for better IDE integration and (b) fixes a potential memory leak related to pinning external views into memory already tracked by our interface. It adds several tests for this memory fix

Three steps to a uv installation

Assuming we already have an Axom installation in ${AXOM_INSTALL}

   # 1. create a uv virtual environment
   $ uv venv --python $(which python3)

   # 2. build the wheel against an Axom installation (internal logic adds the Conduit that axom was built against)
   $ uv pip install /path/to/axom/src/python \
       -C cmake.define.AXOM_DIR="$AXOM_INSTALL/lib/cmake"

   # 3. run an axom.sidre Python script through this environment
   $ uv run python -c "import axom.sidre, conduit, numpy; print(axom.__version__)"

Running in Jupyter w/ type hints

   # 1. install Jupyterlab, and optionally, extra language servers
   $ uv pip install jupyterlab ipykernel
   $ uv pip install jupyterlab-lsp 'python-lsp-server[all]'

   # 2. generate an Axom kernel
   $ uv run python -m ipykernel install --user --name axom --display-name "Axom (uv)"

   # 3. run jupyter
   $ uv run jupyter lab

• Our Axom kernel has tab completion on the API:

• It also has tooltips for the API (via shift+tab)

• Somewhat larger example:

Limitations

• Usage errors in Jupyter currently pipe through cerr (via SLIC_WARNING) and don't show up in the error Jupyter output cell. E.g if you run the cell to create a view or group a second time, it will silently get converted to None.
  • We should use exceptions in the Python interface instead of slic errors/warnings. I'll address that in a separate PR. See Sidre's Python interface should throw errors instead of silently failing #1938
• I didn't test this with hip or cuda configs. Let's also handle that in a separate PR

[kennyweiss]

This user doc about Python will move out of sidre once we have bindings for a second component

[kennyweiss]

This block contains fixes for pinning memory in blocks we've already pinned

[kennyweiss]

The nb::arg() additions help Jupyter and IDEs get the names of functions (instead of just arg1, arg2

[kennyweiss]

The noconvert() prohibits casts like float -> int. (see block comment below in this file)

[kennyweiss]

This file helps us get the conduit that Axom was configured against

[kennyweiss]

This is the CMake project that uv uses to set up the wheel for uv.

I tried to make it as simple as possible, but there were some necessary complexities.
I'm not sure how thoroughly tests the stable ABI options are, but they seemed worthwhile.

[kennyweiss]

This pyproject.toml file sets up the wheel and its dependencies.