Training LLMs to Reason in a Continuous Latent Space