Augment and improve configuration management and loading

[Lawouach]

The problem

Over the last year or so, some users have indicated wanting more from how they load configurations. Historically, configurations (and secrets) were designed to be declared inside the experiment (to make it fully standalone). But, as you grow the number of experiments or depending how you operate them, this becomes a constraint not an enabler.

The current implementation has grown capabilities but it is now becoming complex to understand and maintain while not covering fully all the needs.

This issue here is to discuss what are the new requirements we must take into account and find the most appropriate approach.

What does the community need/want? why?