Fix Token Classification Integration

[gokceuludogan]

This pull request addresses the integration of the token classification model of Turna encoder and BERTurk.

Changes:

• NERDatasets has been updated to handle label outputs as well as text outputs.
• T5ForSequenceClassification has been renamed to T5ForClassification to include token classification, and the forward function has been updated for token classification.
• TrainerForClassification has been updated to reflect changes in T5ForClassification and AutoModelForTokenClassification.
• DataCollatorForTokenClassification has been added to the trainer for token classification.
• SeqEval has been updated to handle token classification output with -100.
• Added ner_classification_base.py, an implementation of fine-tuning BERTurk on token classification using pure transformers.
• Added ner_classification.py, an implementation of fine-tuning BERTurk on token classification with the transformers trainer, our dataset, and evaluators.

The implementation was tested on BERTurk and mT5-small with the following commands:

python finetune.py --config-name ner_wikiann_token model_name=google/mt5-small
python finetune.py --config-name ner_wikiann_token model_name=dbmdz/bert-base-turkish-cased

[gokceuludogan]

POSDatasets need to be updated as well.