SuperServe: Fine-Grained Inference Serving for Unpredictable Workloads