Beware of Fragmentation: Scheduling GPU-Sharing Workloads with Fragmentation Gradient Descent