Train with system prompt

Nobody has claimed this yet.

When using:

```
torchrun --nproc_per_node=2 --master_port=20001 fastchat/train/train.py \
--model_name_or_path lmsys/vicuna-7b-v1.5 \
--data_path data/dummy_conversation.json \
--bf16 True \
--output_dir output_vicuna \
--num_train_epochs 3 \
--per_device_train_batch_size 2 \
--per_device_eval_batch_size 2 \
--gradient_accumulation_steps 16 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 1200 \
--save_total_limit 10 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--fsdp "full_shard auto_wrap" \
--fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
--tf32 True \
--model_max_length 2048 \
--gradient_checkpointing True \
--lazy_preprocess True
```

The fine-tuning is processed with the following system prompt/message:
"A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions."

It would be interesting to specify the prompt to be used,
This what it would look like, inspired by https://github.com/lm-sys/FastChat/blob/main/data/dummy_conversation.json
no prompt would fallback to default prompt:

```
[
{
"id": "identity_0",
"system_message": "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant name is Vick.",
"conversations": [
{
"from": "human",
"value": "Who are you?"
},
{
"from": "gpt",
"value": "I am Vick, a language model trained by researchers from Large Model Systems Organization (LMSYS)."
},
{
"from": "human",
"value": "Have a nice day!"
},
{
"from": "gpt",
"value": "You too!"
}
]
},
{
"id": "identity_1",
"conversations": [
{
"from": "human",
"value": "Who are you?"
},
{
"from": "gpt",
"value": "My name is Vicuna, and I'm a language model developed by Large Model Systems Organization (LMSYS)."
}
]
}
]
```

I can do a pull request for this, if this is considered an interesting feature

Start in fastchat/train/train.py, using the training command in the issue to trace how conversation data is loaded and how the system message is selected. Compare the requested system_message field with data/dummy_conversation.json and verify that records with it use that value while records without it retain the default prompt.

Written by the indexing model from the issue text.