[Benchmark] Fix problems found in XSTest, M3oralbench, MSSBench, MMSafetyBench, Flames and SIUO_GEN.

[Gugugugugutian]

This pull request adds support for several new datasets to the evaluation framework and aligns their evaluation logic and default judge models with the SafeWork-R1 standards. The most significant changes are the addition of new dataset classes, updates to dataset registration, and modifications to the default judge model selection logic.

Better dataset support and evaluation logic:

• Added new dataset classes: MMSafetyBenchDataset, MSSBenchDataset, SIUODataset, SIUOGenDataset, SIUOMCQDataset, XSTestDataset, FlamesDataset, and M3oralBenchDataset, each with custom prompt construction and evaluation logic to match their respective benchmarks and SafeWork-R1 conventions. [1] [2] [3]

• Better detailed evaluation and scoring for the new datasets, including category-wise and overall metrics, and robust judge logic for both rule-based and model-based scoring. [1] [2] [3]

Judge model selection improvements:

• Updated the default judge model selection logic in run.py to assign the appropriate SafeWork-R1 judge models for the new datasets (gpt-4o-2024-11-20, gpt-4o, and gpt-4o-mini as appropriate).