Pushing the Limits of LLM Quantization via the Linearity Theorem