Attention and Transformer blocks — step 4 of 7

Complete one attention row: dot the query against every key (scaled by √dim), softmax into weights, mix the values.