Thanks for starting on this @ev-br. This is mostly right, but it should use Pixi like we do on SciPy, rather than pull CuPy from PyPI - that is good for both maintenance load (same config) and for stability & performance. CUDA libraries, compilers, etc. all come from conda-forge. And that is also better for the ability to run the CI job locally with the exact same dependencies.

I'll figure out how to connect the admin pieces (credit card & billing).