Can Language Models Induce Grammatical Knowledge from Indirect Evidence?