Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning