Unveiling Intrinsic Text Bias in Multimodal Large Language Models through Attention Key-Space Analysis