DRA Structured Parameters

This KEP originally defined an extension of the "classic" DRA #3063 KEP. Now the roles are reversed: this KEP defines the base functionality and #3063 adds an optional extension. Users are increasingly deploying Kubernetes as management solution for new workloads (batch processing) and in new environments (edge computing). Such workloads no longer need just RAM and CPU, but also access to specialized hardware. With upcoming enhancements of data center interconnects, accelerators can be installed outside of specific nodes and be connected to nodes dynamically as needed. This KEP introduces a new API for describing which of these new resources a pod needs. Typically, such resources are devices like a GPU or other kinds of accelerators. The API supports: - Network-attached devices. The existing device plugin API is limited to hardware on a node. - Sharing of allocated devices between multiple containers or pods. The device plugin API currently cannot share devices at all. It could be extended to share devices between containers in a single pod, but supporting sharing between pods would need a completely new API similar to the one in this KEP. - Using a device that is expensive to initialize multiple times in different pods. This is not possible at the moment. - Custom parameters that describe device configuration. With the current Pod API, annotations have to be used to capture such parameters and then hacks are needed to access them from a CSI driver or device plugin. Support for new hardware will be provided by hardware vendor add-ons. Those add-ons are called "DRA drivers". They are responsible for reporting available devices in a format defined and understood by Kubernetes and for configuring hardware before it is used. Kubernetes handles the allocation of those devices as part of pod schedul