Add ideas from gemma scope to allow you to grok the model behavior

[Shuyib]

https://deepmind.google/blog/gemma-scope-2-helping-the-ai-safety-community-deepen-understanding-of-complex-language-model-behavior/