From Tokens to Words: On the Inner Lexicon of LLMs