ParaAMR: A Large-Scale Syntactically Diverse Paraphrase Dataset by AMR Back-Translation