rm-accel 0.5.0

Found in the ai-rm end-to-end run (ai-rm#32): two training processes started on a node outside Kubernetes held both
of its RNGD cards, the device plugin and the AcceleratorProfiles reported them free, and a decode pod placed there
failed with EBUSY (#9).

Added

• Card occupancy (rm_abstract.airm.occupancy): per card, the holders from furiosa-smi ps --format json, the
  Rebellions driver's ctx_info / rbln-stat --json contexts and /proc/<pid>/fd, each classified by
  /proc/<pid>/cgroup as k8s (pod) or host. Published in spec.extra as inUse and holders (pid, executable
  basename, managed). A card held by a host process also gets health: busy and healthy: "false", which
  rm-controller already skips. AIRM_OCCUPANCY=0 turns it off; without the host PID namespace no occupancy is
  published (with a warning).
• Node label airm.ai-computing.io/host-held=<n>; /metrics series rm_accel_accelerators_in_use and
  rm_accel_accelerators_host_held.
• Chart values hostPID (default true) and occupancy (default true).

Verified on: furiosa-smi 2026.4.0, rbln-stat 3.0.0, Python 3.10/3.11, cgroup v2 + containerd. Read-only dry-run on two NPU nodes: host-held RNGD cards reported busy, idle cards free; rm-controller excludes the busy cards (#10).
Install: chart deploy/helm/rm-abstract-node 0.6.0 (hostPID: true is needed to see host processes), or pip install https://github.com/ai-computing/rm-accel/archive/refs/tags/v0.5.0.tar.gz.
Limits: GPUs not covered; a busy ATOM and k8s-held cards are verified by fixtures only.