Optimizing Memory Mapping for Dataflow Inference Accelerators