[feat.] support encoder-decoder tuning and ChatGLM, Vicuna inference

[shizhediao]

Added new features:

1. encoder-decoder architecture fine-tuning (e.g., T5-based model)
2. ChatGLM inference
3. Vicuna inference

[research4pan]

Thanks! This feature is critical for support of ChatGLM 👍 Current commit may require further implementations in Datasets and handling of "text2text" in finetuner, so the merge is temporarily blocked.