how does deepseek r1's mixture of experts (moe) architecture enhance its performance