Adding support for multi-task training

[ozgurozdemir]

• TASK_INFO and use_textual_output are added for tr_datasets.py
• dataset_processor is modified to accept list of dataset names.
• EvaluatorForConditionalGeneration.compute_metrics is updated for multi-task configuration.

[gokceuludogan]

Thank you very much for your significant contributions. I have a few suggestions:

• Instead of adding use_textual_output to each dataset individually, we should consider implementing it directly in BaseDataset since it's a common feature across all datasets.
• The term use_textual_output might not be clear. Perhaps renaming it to task_prefix could provide more clarity.
• For allowing testing, it would be great to add configuration options for multitask training in conf and refactor the existing finetune.py and eval.py scripts or consider implementing new ones.
• Finally, adding a tutorial on multitask training in our documentation would greatly benefit our users.