Remove use of MemoizeJac

[nrontsis]

Removes use of scipy.optimize.optimize.MemoizeJac because it is not publicly exposed by recent versions of scipy. See #160 for more details. This is done by dropping support for passing jac=True in the scipy interface.

Testing: The jax-scipy example was verified to work.

[moorepants]

I looked a bit more closely here. See this test:

https://github.com/mechmotum/cyipopt/blob/master/cyipopt/tests/unit/test_scipy_optional.py#L164

If you provide a single function that calculates the objective and its gradient (similarly for the constraints), then you set jac=True and the function is cached so that the function and gradient are never calculated twice (because they are called at least twice for the same args).

The change you have here probably loses that functionality. This isn't well documented, but I think it is correct functionality to use MemoizeJac to avoid the duplicate computations.

[nrontsis]

Thanks, I agree that if a combined callback is used, then using some of sort of caching makes sense.

But do you think it’s worth maintaining support for combined callbacks at the expense of using an effectively nonpublic scipy import?

[moorepants]

I've opened a fix in scipy. The class was public at one point and we can see if they'll return it to being public.

[moorepants]

Thanks for the submission. I'm closing in favor of #183