GuidedVLA: Specifying Task-Relevant Factors via Plug-and-Play Action Attention Specialization