Inference stop_str missing filter fix

[Trangle]

Why are these changes needed?

1. Fro stop_str = ["</s>", "<|im_end|>"], although "<| im_ end |>" has been added to "special_tokens", there some cases like

I am a chatbot created by researchers.<
I am a chatbot created by researchers.<|
I am a chatbot created by researchers.<|im_
I am a chatbot created by researchers.<|im_end
I am a chatbot created by researchers.<|im_end|><
I am a chatbot created by researchers.<|im_end|><|im_end|><

The original logic will only find the last<| im_end|> through rfind, which in turn leads to filtering failure.
2. When no stop is generated and "rfind_start" is not updated, the starting position is always found, which is time-consuming.
3. Determine the search subset based on the maximum "stop_token" length to improve search efficiency.

Related issue number (if applicable)

Checks

• I've run format.sh to lint the changes in this PR.
• I've included any doc changes needed.
• I've made sure the relevant tests are passing (if applicable).

[Trangle]

@merrymercy Help deal this?