Models In a Spelling Bee: Language Models Implicitly Learn the Character Composition of Tokens