Self-Exploring Language Models: Active Preference Elicitation for Online Alignment