Linear Neural Parsing and Hybrid Enhancement for Universal Dependencies