trsm / trmm-like operation optimisations

[willtebbutt]

The current implementations for backsolving in generic.jl and trmm / trsm in blas.jl are sub-optimal. Certain quantities that do not need to be recomputed are being recomputed. In the situation that the gradient w.r.t. both (matrix-valued) arguments is requested, the recomputed quantity is actually the sensitivity w.r.t. one of the arguments, thus this optimisation is strictly better than the current implementation in such scenarios.

The solution to this is conceptually straightforward, and we even have part of the infrastructure required to make this work sorted (the currently unused third argument to the sensitivities).