Make pg_gpu pip-installable from PyPI

[andrewkern]

Closes #115.

The easiest fix here is to copy our pixi env info (pixi.toml) to pyproject.toml. Currently pyproject.toml declared no dependencies at all -- every runtime requirement lived only in pixi.toml (by design), so a pip install produced a package that couldn't import.

Changes

• pyproject.toml: declare the full runtime stack in [project.dependencies], with GPU support (CuPy, kvikio, nvcomp) as hard requirements rather than optional extras. Add [project.optional-dependencies] for dev / docs / moments.
  • Uses cupy-cuda12x[ctk] so the CUDA toolkit headers CuPy needs to JIT-compile its kernels are pulled from PyPI
  • Adds matplotlib, seaborn (imported by the plotting module via init) and tskit (used directly by HaplotypeMatrix)
  • Fixes the Homepage URL to the kr-colab org.
• pixi.toml: pyproject is now the single source of truth for the PyPI/GPU stack; the duplicated deps are removed and come transitively via the editable install. Dropped the NVIDIA extra-index-url (kvikio-cu12 / nvidia-nvcomp-cu12 are now on the default PyPI index). Added conda cuda-cudart-dev so CuPy's NVRTC compilation finds toolkit headers inside the pixi/conda environment too.
• Docs: README and docs/source/installation.rst gain an "Installation into an Existing Environment" section; pixi remains the recommended
  path.

Install paths

  # Recommended (pinned, reproducible)
  pixi install && pixi shell

  # Into an existing conda/venv environment
  pip install "git+https://github.com/kr-colab/pg_gpu.git"

[andrewkern]

@mufernando will you try this branch out on your setup?

[nspope]

Why are we pinned to python <3.13? This is going to be annoying when integrating pg_gpu into existing pipelines/environments. Is it because of CUDA toolchain or kvikio? Glancing through the lock file, it's not obvious to me where the constraint is.

[andrewkern]

ah good catch. we're not in capping the install in pyproject.toml. this is vestigal and i'll lift it.

[andrewkern]

also -- i'm 95% confident everything (including the cuda stack) installs via pip cleanly, so we can publish to pypi

[nspope]

installs and imports fine for me with

python3 -m pip install "git+https://github.com/kr-colab/pg_gpu.git@feature/pip-installable-deps"
python3 -c "import pg_gpu"

(whereas main branch fails as expected on import)

[andrewkern]

excellent. when we merge this we can publish to pypi

[mufernando]

sorry for the delay! it works for me as well. thanks, @andrewkern!