DynamicSupervisor holding one MLServe.ModelInstance per loaded {name, version}.

Models arrive and leave at runtime — from configuration at boot, from MLServe.load_model/2 during a deploy, from a version promotion — which is exactly what a dynamic supervisor is for. Because each child is a whole subtree, two versions of the same model coexist as independent siblings, which is what makes canary rollout and drain-on-unload possible.