Fix/audit corrections

[alirezabayatmk]

Summary 📝

A few steps could give wrong or degraded results on some models; this fixes them.

Details

1. Model defaults dtype="auto": float32 for small models and bfloat16 for large ones, so small-model logits stay correct.
2. The Intervene step now gives an error when a direction targets no hooked site, instead of returning the clean generation as the intervened one.
3. LogitLens no longer stores the full vocab probabilities by default; pass store_full_probs=True to keep them.
4. LogitAttribution now handles Gemma-family RMSNorm, so its completeness holds there.

Checks

• Tested Changes: added tests for weight ablation, SAE capture, the logit-lens ground truth, and the dtype estimate; validated on gpt2, pythia-160m, Llama-3.2-1B, Qwen2.5-0.5B, and gemma-2-2b.