"Flex Tape Can't Fix That": Bias and Misinformation in Edited Language Models