In-Context Learning of a Linear Transformer Block: Benefits of the MLP Component and One-Step GD Initialization