CPU-only LLM inference engine in pure Rust — 4-bit quantized models, hand-written AVX2 k…