[Ready] align sft formats & new ops

[drcege]

Summary

• Unified SFT data format
  • add query_key, response_key, history_key in base_op
• New OPs & unit-tests
  • generate_qa_from_examples_mapper
  • generate_qa_from_text_mapper
  • optimize_qa_mapper
  • optimize_query_mapper
  • optimize_response_mapper
• Fixed model inference (use chat_template)
  • use chat interface for vllm
  • use pipeline for HF
• Updated docs
• Optimized model loading
  • load model directly to GPU instead of moving it to CPU first
  • properly configure vllm to load only in tensor parallel mode
  • preload only models without file locks
  • sort prepare_*_model functions

[drcege]

Note: extract_qa_mapper 改为 generate_qa_from_text_mapper 后输出格式变化，不再产出 chatml 的 json string，而是分解为具体的 sft 样本，注意后续对齐。

[BeachWang]

LGTM