Fluid Batching: Exit-Aware Preemptive Serving of Early-Exit Neural Networks on Edge NPUs.