Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs