[misc] fix new tokens adding

[flashJd]

What does this PR do?

Fixes #7119
Fixes #7080

Background:
When sft distilling data from deepseek-R1, we want to add two new tokens and to vocab
but after add two configs: new_special_tokens: and and resize_vocab: true
then do sft, we still can't get and in vllm reference

Cause:
1).I debugged the code and found the config new_special_tokens will call tokenizer.add_special_tokens which added new token to additional_special_tokens
2)vllm will default skip special tokens thus we can't get and in vllm reference

Fix:
In this backgroud, we should just add new tokens to vocab, not make them to be special tokens(special tokens like <|im_end|> always has special function), so just add a config: new_normal_tokens which call tokenizer.add_tokens, after that new tokens will be displayed in reference

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests?