Efficient Adaptive Activation Rounding for Post-Training Quantization