this PR deprecates the only_last_token parameter in favor of an index which can be much more generic. The index can be one of an int, a tensor, or None where an int will gather the same index (token) for each sequence in the batch, a tensor will gather potentially different index's (token) for each sequence in the batch, and None will return the whole sequence. This was primarily added in the case a user uses right padding, where the last token exist somewhere in the middle of the sequence.