Task-Stratified Knowledge Scaling Laws for Post-Training Quantized Large Language Models