Train Once, Deploy Anywhere: Matryoshka Representation Learning for Multimodal Recommendation F i n d i n g s - E M N L P