OnSRAM: Efficient Inter-Node On-Chip Scratchpad Management in Deep Learning Accelerators