Multi-Head Self-Attention

One linear projection, three views, H parallel heads, one mask. The attention block as the model actually uses it.