Powering cost-efficient AI inference at scale with Cloud TPU v5e on GKE