Implement backend for FiniteDifferences

[sethaxen]

This PR implements a backend for FiniteDifferences.jl. It basically combines the 3 declarations in the tests into 1, with generalization of pullback_function to multi-output functions.

Informal benchmarks showed that all three of these methods were necessary for speed. They also showed that with these defined, our default gradient and multi-arg derivative methods are faster than custom overloads using jvp and grad. A single-arg derivative method was faster than our default, yet for some reason it was less accurate.

[codecov-commenter]

Codecov Report

Merging #37 (4123a74) into master (979fa04) will increase coverage by 0.58%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #37      +/-   ##
==========================================
+ Coverage   80.05%   80.63%   +0.58%     
==========================================
  Files           2        3       +1     
  Lines         396      408      +12     
==========================================
+ Hits          317      329      +12     
  Misses         79       79              

Impacted Files	Coverage Δ
src/AbstractDifferentiation.jl	78.27% <100.00%> (+0.06%)	⬆️
src/finitedifferences.jl	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 979fa04...4123a74. Read the comment docs.

[mohdibntarek]

thanks Seth!