AVIS: Autonomous Visual Information Seeking with Large Language Model Agent