Add optional cgroup-direct process attribution for non-MIG shared GPUs

[SMS-Derfflinger]

Is this a new feature, an improvement, or a change to existing functionality?

Improvement

Please provide a clear description of the problem this feature solves

#594 added per-process metrics for shared GPUs by selecting candidate pods from kubelet PodResources and then mapping NVML PIDs to those pods through cgroups.
Our some non-MIG MPS device plugins expose logical device IDs such as:

GPU-6a16b9c2-...::5
GPU-6a16b9c2-...::19

Our shared-GPU device plugin exposes 48 logical slots per physical GPU. A workload pod usually requests 6 or 7 slots, while its processes run on one physical GPU.
These logical assignments are allocation records and may not reliably describe runtime physical GPU ownership. They can also cause the same pod-level FB_USED value to be emitted once per logical slot:

DCGM_FI_DEV_FB_USED{UUID="GPU-6a16b9c2-...",pod_uid="1843d150-...",vgpu="5",...} 10056
DCGM_FI_DEV_FB_USED{UUID="GPU-6a16b9c2-...",pod_uid="1843d150-...",vgpu="19",...} 10056

This request only covers non-MIG DCGM_FI_DEV_FB_USED.

Feature Description

As an operator using a MPS(Multi-Process Service) shared-GPU device plugin, I want an optional process attribution mode based directly on NVML PIDs and cgroups, so that pod-level GPU memory reflects actual runtime ownership instead of logical slot assignments.

Describe your ideal solution

Add an optional cgroup-direct attribution mode:

physical GPU UUID -> NVML PID -> /proc/<pid>/cgroup -> pod UID

Aggregate DCGM_FI_DEV_FB_USED by:

physical GPU UUID + pod UID

The expected pod-level output is:

DCGM_FI_DEV_FB_USED{UUID="GPU-6a16b9c2-...",pod_uid="1843d150-...",pod="workload-1",namespace="default",...} 10056

PodResources and the Kubernetes informer should not determine physical GPU ownership in this mode. The existing PodResources-based mode should remain the default. Non-FB_USED metrics should remain unchanged.

Additional context

Tested with:

dcgm-exporter: 4.5.3-4.8.2
GPU: NVIDIA RTX 6000D
Driver: 580.126.09
Environment: Kubernetes with an MPS shared-GPU device plugin

This is related to #587 and #594. I can submit a pull request if this approach is acceptable.