toy-attention-residuals

added prenorms to mha and mlp layers, aded back toggle for standard residual connections, added sinusoidal positional embeddings, training loop refactor

attnres works!

fixed mha, added another norm, next step to fully plumb in attnRes. also realized paper treats each selfattn or mlp as an individual layer. means in my fullattnreslayer class i need 2 query weight vecs

added mha

layer 1 gradients increase after using attnres compared to w/o. going to add positional enc later to see how that changes stuff. see how that affects loss.