amdgpu_job_dependency busy-loops forever on amdgpu_vmid_grab allocation failure, wedging the ring scheduler kthread (unbounded spin + DRM_ERROR flood, ring-wide GPU DoS)

Summary

The while-loop's three exits are all under vmid_grab's control; its alloc-failure paths (-ENOMEM at 4 sites) return without setting vmid or adding to sync, so the sched worker spins at 100% with unbounded DRM_ERROR flood, stalling all CS on the ring. Unpriv-triggerable with GPU access racing kmem exhaustion; upstream-identical shape. Hardware-gated on guest (no AMD VGA - verified). Break-on-persistent-error fix (lets ib_schedule reject cleanly).