Add example of integration with vLLM

[rlouf]

This example currently does not work with multiple prompts, because self.fsm_state will be updated every time self.__call__ is called. So with several prompts, self.fsm_states will be updated, at each step, as many times as there are sequences. This can be avoided by having self.fsm_states as a DefaultDict and passing the seq_id to logits_processor.

Looking at the code, it might be a good idea to revert back to the original tokenizer interface.