v0.2.3

Align the detect path with gpustack-operator's device manager, split the
information query from the usage query, and report why a workload stopped.

- Nine vendor detectors audited against the operator; MetaX no longer drops
  every card on a virtualization-enabled host, MThreads no longer crashes the
  whole pass, Ascend detects on older drivers, Iluvatar falls back to v1
  memory, and Cambricon is a real CNDev binding instead of a cnmon shell-out.
- detect_info() and detect_usage() are both abstract; detect_devices(usage=False)
  skips every metric call.
- Whole-card reporting: the vGPU classification is gone for every vendor, and
  MIG instances live in appendix["mig_devices"].
- Device.index is the enumeration index; driver-physical numbering moved to the
  appendix, and the physical-index switch is retired. The deployer pins
  CUDA_DEVICE_ORDER=PCI_BUS_ID for containers given every device.
- WorkloadStatus.exits reports per-container exit code, reason and timestamps on
  Docker, Podman and Kubernetes, and a Kubernetes workload whose image cannot be
  pulled reports Failed with the Pod Event appended.

Deployments consuming the Kubernetes deployer need one new RBAC rule:
get/list on core events.

Closes the runtime side of gpustack/gpustack#6041, #5869 and #4217.