Einsum

[petarmhg]

Gradient calculations up to: dL / ds_t * ... * ds_{bp_lim} / d_
are accurate (_ being a standing for z, r, and h)

The einsums passed to W_.backward() and U_.backward() to calculate d_ / W_ and d_ / U_
however, do not correctly calculate the derivatives and sum for the total loss with respect to the weights.