[Refactor] Refactor the preprocess of dataset

[LZHgrla]

dataset_map_fn: map the original data to the format of

[{
    "system": "You are an assistant.",
    "input": "Tell me a story.",
    "output": "One day a student went to schoool."
}]

template_map_fn: [defined on PROMPT_TEMPLATE] map the above format data to trainable data:

[{
    "input": "<|System|>: You are an assistant.\n<|User|>: Tell me a story.<eoh>\n<|Bot|>:",
    "output": "One day a student went to schoool."
}]