Now the CI relies on pixi

This MR switches the CI from a combination of mamba install and pip install to pixi.

Notes about the Docker image:

• As suggested by Ernestas, I have tried the official pixi Docker image, ghcr.io/prefix-dev/pixi, but:
  • It does not come with git, and it takes longer to install git than pixi.
  • It does not support Windows. Ernestas wrote that for TopFarm they created registry.windenergy.dtu.dk/dockerimages/windows-pixi:ltsc2019. However, that will need to be updated at every new release of pixi.
• I have tried ubuntu:latest, but:
  • It does not come with git nor curl, and it takes a lot of time to install them.

In conclusion, we stay with condaforge/miniforge3:latest even if pixi will install python anyway.

My goal was to replicate the former CI as closely as possible. However, while doing so I noticed that build_pages requires both the docs and the test dependencies. We might want to clean this up later.

For safety reasons, I have configured pixi to not install any package newer than 1 month. I will add a job with pip-audit later.

The test suite was, and is still based on, python 3.12. We might also try python 3.13, since that is the newest version supported by TensorFlow.