Glitch Tokens in Large Language Models: Categorization Taxonomy and Effective Detection