Exact Hard Monotonic Attention for Character-Level Transduction