ExLM: Rethinking the Impact of [MASK] Tokens in Masked Language Models