Part 1: Simplify the existing API

[nzarif]

This part includes 3 tasks:

• Set the model to the inference mode by default + ensure training script is not impacted + update failing unit-tests

• Remove the hf-format from Next Item Prediction task + ensure training / inference are not impacted + update failing unit-tests

• Replace ignore_masking with testing flag + ensure training / inference are not impacted + update failing unit-tests

P.S: This PR removes the hf_format only for NextItemPredictionTask. Other tasks such as BinaryClassificationTask and RegressionTask still use hf_format as removing the flag for them at this stage does not make sense for them, they use the PredictionTask.forward() only for inference and use compute_loss() for training and eval.