SQuALITY: Building a Long-Document Summarization Dataset the Hard Way