custom c++ cuda kernel for the modified self-attention

[RuiFilipeCampos]

this is actually critical to get the performance out of it, without a custom implementation the layer is actually slower than the scaled dot product