From Form(s) to Meaning: Probing the Semantic Depths of Language Models Using Multisense Consistency