Self-Regularization with Latent Space Explanations for Controllable LLM-based Classification