Rework chapter about training on k8s pod

[rmarquis]

Since the model training on a k8s pod using CML runner registration has proved to be problematic - seemingly due to an incompatibility with CML and recent version of Kubernetes, see #126) and that CML is now pretty much in maintenance mode, we won't use CML to instantiate the training on a k8s pod.

Instead, we'll use 2 usual runners, as prototyped in the SwissImage satellite image segmentation project.

Ensure that k8s training is documented for both GitHub actions and (ideally) GitLab CI/CD.