fix evaluation during training for t5

[lukaszett]

This fixes #1524: When evaluating a T5 model during training a TypeError is thrown.

This was because the eval_model method tried to get the inputs for the evaluation from the wrong dataset. I think eval_dataset contains the encoded inputs (and does not have column names) - the fix was just replacing the reference to eval_dataset with eval_data which is the dataframe that eval_data gets from the arguments to the train_model method.

There might be some superfluous work here as the model feeds the inputs through the model twice instead of just doing additional processing of the outputs of T5model.evaluate() - but my solution works as a hotfix.