Update docker environment, use shared runners, and plafrim for cuda/rocm

Use new docker images based on Ubuntu 24.04 : base for testing analysis for static and dynamic analysis.

Use the gitlab's instance runners dedicated to docker jobs instead of our old custom virtual machines from ci.inria.fr which will be destroyed in a near future.

The shared runners do not accept docker images larger than 10 GB (which is reasonnable). As a result we must stop using the big docker image with the union of all solverstack's dependencies, and prefer using a basic image plus some precise dependencies project by project. cuda and rocm stacks are very heavy (30 GB) and are specific HPC components which are especially used on supercomputers, so let's validate cuda and rocm builds on plafrim supercomputer, and using GNU Guix to also check that latest chameleon is working in guix (to kill two birds with one stone).

Clean some CI scripts.