Barking Up The Syntactic Tree: Enhancing VLM Training with Syntactic Losses