Refactor streaming and g-sampling FSM logics

[jeffreymeetkai]

• prompt_template.initialize_fsm_gen_state initializes prompt template's FSM gen state. Both streaming and gsampling will call this method at the start.
• In streaming, the generator will repeatedly call prompt_template.stream_delta_text method that streams specific delta texts and updates the gen state at every iteration
• In gsampling, monkey-patched async_llm_engine will call prompt_template.grammar_sample method that grammar samples the tokens and updates the gen state at every iteration
• prompt_template.update_fsm_gen_state updates the gen_state.
• If both gsampling is enabled and the request is a streaming request, grammar sampling will first create and maintain a FSM to produce the grammar-sampled token to yield to the wrapping generator. The wrapping streaming generator will also create and maintain a separate FSM to stream the grammar-sampled tokens.

[khai-meetkai]

Oh why this is always True?

[jeffreymeetkai]

If this is True, it means we need to stream an empty chunk before streaming the chunks containing the function name and arguments. Thereafter, we will set this to False.

[khai-meetkai]

I think we can create a function for getting options, this is kind of duplicate in function: stream_delta_text and grammar_sample

[jeffreymeetkai]

Sure will do