Support RDMA NIC sharing

[gauravkghildiyal]

We want to support sharing a single RDMA NIC across multiple pods

Context: (Based on a discussion on slack) There are use cases within LLM inference which require scheduling multiple pods on a single Node which consume accelerator (GPU/TPU). The Node itself may only have one RDMA device, which means for multi-node inference, the RDMA device would need to be shared between all pods on the single node to allow RDMA traffic to other nodes.

E.g.

CPU / PCIe Root Complex 0
 ├── GPU0
 ├── GPU1
 ├── GPU2
 ├── GPU3
 └── NIC0

Possible solution:

1. Create IPVLANs for each pod requesting an RDMA NIC to share the underlying RDMA device.
2. Use the Consumable Capacity (kep.k8s.io/5075) in DRA to model this resource sharing capability.