Text this: PKTQ: Per‐Kernel Thresholded Quantization With Size‐Constrained Two‐Stage Optimization for CNNs