Fix exponentiation for NaNMath.pow

[jClugstor]

Currently an example like this:

using ForwardDiff
using NaNMath

function new_pow(x)
    NaNMath.pow(x[1],x[2])
end

ForwardDiff.gradient(new_pow, [-1.0, 2.0])

errors. I'd like to make it so that it returns a NaN instead.

Fixes #716

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 86.75%. Comparing base (c310fb5) to head (db4d9f9).
Report is 11 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #717      +/-   ##
==========================================
- Coverage   89.57%   86.75%   -2.83%     
==========================================
  Files          11       10       -1     
  Lines         969      921      -48     
==========================================
- Hits          868      799      -69     
- Misses        101      122      +21     

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

[mcabbott]

Can you add a few cases to the tests? Ideally NaN/not/negative through both paths.