Memory-efficient Patch-based Inference for Tiny Deep Learning