We have a modest on-prem GPU cluster shared across training and inference experiments, and scheduling fights are constant. Slurm works but feels heavyweight for our size. What are small teams using to share GPUs sanely — Kubernetes with GPU operators, RunPod-style, something else?