Assessing Large Language Models for Online Extremism Research: Identification, Explanation, and New Knowledge