Enhance model service to reduce fragmentation

[asm582]

ModelService creates compute and non-compute resources. Under compute resources, decode and prefill deployment can be made, where each can have multiple replicas. The default Scheduler will spread such pods. We need a mechanism to pack pods on a node. Here is a scenario: Assume the cluster has two nodes, each having two GPUs. If prefill and decode deployment with one replica places a pod on both the nodes, this will cause the next workload replica requesting two GPUs per pod not to start, even if the cluster has two GPUs.