[fix] pass@k and self-consistency

[huyiwen]

• set default temperature to 0 for vLLM backend 同时测试多个子集的结果与单独测试每个子集的结果不同 #275
• fix pass@k and self-consistency

The original aggregate method of pass@k and self-consistency does not apply to dataset with multiple subsets. So now we apply aggregate for each subset separately.