On the Compressibility of Quantized Large Language Models