CASPER: Inferring Diverse Intents for Assistive Teleoperation with Vision Language Models