[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