Don't Deceive Me: Mitigating Gaslighting through Attention Reallocation in LMMs