dynamic resource allocation

Originally, this KEP introduced DRA in Kubernetes 1.26 and the "structured parameters" KEP added an extension. Now the roles are reversed: #4381 defines the base functionality and this KEP is an optional extension. In Kubernetes 1.32, this KEP has been withdrawn and all code related to it gets removed. #4381 continues. The main objections against this KEP that led to this decision were: - Lack of support for cluster autoscaling because a cluster autoscaler cannot reason about resource availability when adding or removing nodes. - Complex back-and-forth through the apiserver while scheduler and DRA drivers negotiate how to allocate a ResourceClaim. With #4381, DRA drivers are limited by what the structured parameter model(s) defined by Kubernetes support. New requirements for future hardware may depend on changing Kubernetes first. With this KEP, parameters and resource availability are completely opaque to Kubernetes. During scheduling of a pod, the kube-scheduler and any DRA driver controller(s) handling claims for the pod communicate back-and-forth through the apiserver by updating a PodSchedulingContext object, ultimately leading to the allocation of all pending claims and the pod being scheduled onto a node. Beware that this approach poses a problem for the Cluster Autoscaler (CA) or for any higher level controller that needs to make decisions for a group of pods (e.g. a job scheduler). It cannot simulate the effect of allocating or deallocating claims over time. Only the third-party DRA drivers have the information available to do this. Structured parameters from #4381 should be used when cluster autoscaling is needed.