Enhance conv prompt and train

[Trangle]

Why are these changes needed?

1. The process of constructing prompts in the conversation has been separated, allowing for more general circumstances such as when the GPT speaks first or when the human speaks multiple times.
2. A prompt generator and manager have been constructed, with each generator handling its own training token_ids. This makes definitions more clear and concatenation more universal.
3. An additional 'train_general' has been introduced, providing support for more general SFT models.

I've already validated the code logic several times and haven't found any potential hidden issues, please help me test it further.

Related issue number (if applicable)

Checks

• I've run format.sh to lint the changes in this PR.
• I've included any doc changes needed.
• I've made sure the relevant tests are passing (if applicable).