Accept that NaN ≈ NaN in test_approx

[oxinabox]

I think this is the behavior we want.
If you get a actual value of NaN and you were expecting a NaN then all seems well.
If both the AD rule and the finite differencing take you here all is well.
(not ok if only one does)

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.42%. Comparing base (683e023) to head (447c9e9).
⚠️ Report is 77 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #220      +/-   ##
==========================================
- Coverage   90.75%   90.42%   -0.34%     
==========================================
  Files          11       11              
  Lines         303      303              
==========================================
- Hits          275      274       -1     
- Misses         28       29       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[oxinabox]

This is not quite the same as JuliaDiff/FiniteDifferences.jl#192
which woulderror if the input is not finite.
Here we are allow output to be NaN,
so we could do both

[mzgubic]

needs a version bump and some tests?

[oxinabox]

idk if this is a good idea or not

[mzgubic]

The only thing I can think of going wrong is getting a NaN for the wrong reason. But seems relatively unlikely (and harmless in the end, at least at the tested point?)