Add MLX 2-bit quantized Qwen3.5-27B with RotorQuant KV cache

initial commit