Unified Training scheduler incorrectly use allocatable for free GPU resources

[Fizzbb]

With Nvidia GPU device plugin, the allocatable GPU is a fixed number, i.e., the number of GPUs plugged to the node. It won't update based on current usage. Therefore, when scheduler decide the target replica (gpu number) for the unified job, the available GPU needs to be calculated by listing existing job's gpu request, also should add some safe exit process (e.g., reduce replica size after fail scheduling N times, drop job after M times of failure) to avoid keeping create/delete unsuccessful pods.

[Fizzbb]

when describe node, at the end it shows allocated resources, which is the correct number. could call describe node command and extract that field. After search the describe code, they also implemented by scanning all the pods on the node. So could just use the old method to scan all the running pods and their requested GPU

[Fizzbb]

Merged with pull request #90 .
Interesting thing about select running pods from a given node is you need to define the Index first with manager. the List MatchingField is not automatically available. Also don't use real path, may shallow with existing ones and cause irrelevant errors.
Kuberbuilder's jobOwnerKey is a reference.
kubernetes-sigs/kubebuilder#1422

[Fizzbb]

add one more reference for Kubernetes Field Selector, https://kubernetes.io/docs/concepts/overview/working-with-objects/field-selectors/
Supported field selectors vary by Kubernetes resource type. All resource types support the metadata.name and metadata.namespace fields. Using unsupported field selectors produces an error.