Investigate scipy trust region optimisation #38
Description
Previously, there were 3 optimisation options, trust region, BFGS, and L-BFGS-B. The trust region method used the package optimistix and used BFGS as the optimisation method, while the others are from scipy. Jan suggested removing optimistix, so I looked into replacing the trust region with a scipy version here. However, none of them seem to be exactly what was used in optimisitix, and seem to compute (or use) the Hessian which we want to avoid because it is expensive. I also didn't know if they had any requirements about the convexity of the problem (and I believe for mixed logit the simulated log likelihood is not convex), so I just removed the trust-region option in PR #35 and issue #24.
In future maybe we could look into a adding one of these if it were to give a better result, and signal to the user that it would be faster/slower and more/less accurate.