improve robustness in case of ipopt failures

we can borrow an idea from the conic Pajarito: adding gradient cuts at OA (MIP) solutions when the subproblem solver fails. this complicates the algorithm somewhat, for example, you want to check each OA solution for near-feasibility in which case it could be a new incumbent.

see DaChoppa.jl, which does not use a subproblem solver. it fails when the initial mixed-integer linear relaxation is unbounded. this can generally be fixed by the user by imposing some initial outer approximation constraints.