CrAM: Credibility-Aware Attention Modification in LLMs for Combating Misinformation in RAG