KLQ: training-free rotation quantization beats SpinQuant at W4A4KV4

KLQ beats all training-free rotation-based quantization methods on W4A4KV4-bits, with Llama 3.2 1B KLQ-quantized outperforming SpinQuant and approaching ReSpinQuant without GPTQ/LDLQ rounding. It is a solo summer research project; the repo covers methods, findings on quantization and geometry, limitations, and proposed experiments.