Gradient-Based Post-Training Quantization: Challenging the Status Quo