KQ-SVD: Compressing the KV Cache with Provable Guarantees on Attention Fidelity