feat: implements on topic guardrail

[ErezSha]

implements an on_topic guardrail for the graph

[asaf]

@ErezSha Is this another LLM call and going to increase latency? we'r already at > 20s

[ErezSha]

That's the only way to do it reliably. It calls gpt-4o, so it's faster.
Also, it's happening asynchronously, in parallel, with the other call entry point makes, so it's doesn't increase latency.