MOI backend inefficiencies

[baggepinnen]

The MOI backend has a number of performance problems

• findnz is called for each evaluation of hessians, this allocates
• a lot of slow indexing of sparse arrays rather than just filling the H.nzval.

In general, the approach taken in MOI appears better in general, they only use vectors of tuples to indicate sparsity structure when the problem is set up, after that, they only operate on the vector of non-zero values rather than dealing with sparse matrices.

The function eval_hessian_lagrangian could probably be rewritten to also do this and save all the expensive sparse matrix operations