The Shaped Transformer: Attention Models in the Infinite Depth-and-Width Limit