LUMIA: Linear probing for Unimodal and MultiModal Membership Inference Attacks leveraging internal LLM states