Accelerating Neural Network Inference with Processing-in-DRAM: From the Edge to the Cloud