Your Language Model Can Handle Non-Canonical Tokenizations