KnowThyself: An Agentic Assistant for LLM Interpretability