[Feat] Multiple datasets evaluation

[huyiwen]

• Support multiple datasets evaluation. Example: python inference.py -m davinci-002 -d hellaswag copa
• Support evaluate category of datasets. Example: python inference.py -m davinci-002 -d mmlu:[stem],[humanities]
• Support AGIEval as a mixture of GenerationDataset and MultipleChoiceDataset. Example python inference.py -m davinci-002 -d agieval
  • Use DatasetCollectionBatchSampler to manage the model initialization context
• Rename enum and task names to clarify
• Improve code styles