A Trainer subclass for Decoder-Only LM with generation in evaluate()

### Feature request The main feature request involves a New Trainer Subclass, similar to Seq2SeqTrainer, but suitable for Decoder-Only LM. ### Motivation `Seq2SeqTrainer` provides a great abstraction for Encoder-Decoder LM, when we need to conduct generation during `evaluate()` But the current implementation of both `Trainer` and `Seq2SeqTrainer` seems to be not suitable for Decoder-Only LM due to the difference of `input_ids` and `labels` between teacher-forcing training and generation-involved evaluation. For example in instruction tuning: - During training (teacher-forcing) ```python input_ids = 'Translation the following texts: {Text in Chinese...} {Text in English...}' labels = 'Translation the following texts: {Text in Chinese...} {Text in English...}' ``` - During evaluation ```python input_ids = 'Translation the following texts: {Text in Chinese...}' labels = '{Text in English...}' ``` So we need to prepare two kinds of inputs_ids during evaluation for calculation of both `loss` and `bleu_metrics`. It leads to different columns in eval_dataset. However, `Trainer._remove_unused_columns()` will remove columns for both `eval_dataset` and `train_dataset` not accepted by `model.forward()`. During training, this behaviour is expected (we only need the teacher-forcing inputs). But it will make evaluation difficult. This feature is nearly identical across all CausalLM models when performing generation during evaluation, making it highly reusable. Given the increasing number of Decoder-only LMs (CausalLMs) in the community, I strongly recommend implementing a dedicated CausalTrainer to simplify deployments. I may have missed something. If there is already a simpler way to customize such a Trainer, please let me know. ### Your contribution I'm willing to help submit a PR. But I'm not familiar with some integrations such as fsdp and deepspeed. I may need someone to help me finish this feature.