MAINT: optimize.minimize: revert gh-11712 to fix macos_arm64_test

[mdhaber]

Reference issue

gh-11649
gh-11712

What does this implement/fix?

macos_arm64_test has a failure in main:

FAILED scipy/optimize/tests/test_minimize_constrained.py::test_gh11649 - UserWarning: delta_grad == 0.0. Check if the approximated function is linear. If the function is linear better results can be obtained by defining the Hessian as zero instead of using quasi-Newton approximations.

This was just a warning, but it exposed a bug introduced in gh-11712. This PR reverts gh-11712 but leaves the test (and strengthens it), xfailing it with strict=True. The original issue, gh-11649, has been re-opened, and the xfail mark can be removed when it is resolved.

Additional information

@dpoerio I know you are not able to merge this, but would you review it since you're familiar with the issue? I think that will help another maintainer.

[ev-br]

LGTM as a stop-gap. When this is merged, please list it in gh-17278

[dpoerio]

Effectively undoes the previous PR (#11712) as intended. I am not currently certain if all of the assertions in the test case are valid (i.e., if the result should be the same as SLSQP algorithm) but that could be revisited once a fix is in place if needed.

[mdhaber]

The minimal objective function values should be close if they end up in the same local minimum. I understand that if there are multiple local minima they might not, but it is a reasonable smoke test to see if trust-constr is misbehaving. Currently, trust-constr ends up with a lower function value than SLSQP, which prompted me to check the constraints (and the equality constraint is violated).

[WarrenWeckesser]

Looks good to me. If no one else has jumped in to merge this sooner, and there are no new comments, I'll merge this on Sunday.