[Bug] ChatGLM2 tokenizer dismatch on `eos_token_id` and `eos_token`

[LZHgrla]

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm2-6b", trust_remote_code=True)
tokenizer.encode(tokenizer.eos_token)  # [64790, 64792, 2893, 30917, 30994]
tokenizer.eos_token_id  # 2

XTuner uses tokenizer.encode(tokenizer.eos_token) instead of tokenizer.eos_token_id to process data. This causes that the fine-tuned ChatGLM2 cannot stop the generation.

[LZHgrla]

Fixed by #107