Exposing Attention Glitches with Flip-Flop Language Modeling