Gemma Scope 2: helping the AI safety community deepen understanding of complex language model behavior