[New Feature] Training support example packing to increase training efficiency

Huggingface trl SFTTrainer has support example packing to increase training efficiency.

SFTTrainer supports example packing, where multiple short examples are packed in the same input sequence to increase training efficiency. This is done with the ConstantLengthDataset utility class that returns constant length chunks of tokens from a stream of examples. To enable the usage of this dataset class, simply pass packing=True to the SFTTrainer constructor.

And, Meta LLAMA2 SFT also use same feature.

For the fine-tuning process, each sample consists of a prompt and an answer. To ensure the model sequence
length is properly filled, we concatenate all the prompts and answers from the training set. A special token is
utilized to separate the prompt and answer segments. We utilize an autoregressive objective and zero-out
the loss on tokens from the user prompt, so as a result, we backpropagate only on answer tokens.

Now, fastchat has already support only compute loss on the assistant outputs. Can further support example packing?

Start with fastchat/train/train.py, especially the existing assistant-output loss handling, and compare it with the linked TRL SFTTrainer packing documentation. Determine how packed sequences and separators should interact with prompt masking; done means training supports packed examples while retaining loss only on assistant outputs.