Evaluating Byte and Wordpiece Level Models for Massively Multilingual Semantic Parsing