Unlocking Data-free Low-bit Quantization with Matrix Decomposition for KV Cache Compression