Word Shape Matters: Robust Machine Translation with Visual Embedding