Shape and Substance: Dual-Layer Side-Channel Attacks on Local Vision-Language Models