Add ability to stop a long running request.

[richard-to]

User has a use case where they have a long running LLM request that may yield a lot of info. If they see the info is going the wrong direction, they want the ability to stop request.

At the moment, I think potentially could be done unintentionally through this bug: #981 (lol).

[wwwillchen]

You can do the following pattern if Mesop websockets is enabled:

@me.stateclass
class State:
  is_cancelled: bool = False

def call_llm_handler():
    for chunk in call_llm():
        if state.is_cancelled:
            raise GeneratorExit
        yield chunk

But without websockets (and a stateful context across multiple concurrent requests), I don't think this is supported.