Implement backend for Tracker

[sethaxen]

Relates #40

[sethaxen]

Tracker's an interesting case because I think you want to avoid calling Tracker.data on the outputs only if you're nesting the AD (note the Hessian failures). So here we may do something like add a type argument to the backend to indicate whether nesting is happening and then lowest would set that to true, and based on that we would unwrap the outputs or not.

[sethaxen]

Actually, even if we track nesting in the constructor, we still have the problem that the test suite uses prod to test Hessians, but Tracker computes the wrong Hessian for prod: FluxML/Tracker.jl#120. Perhaps the testing functions should allow the user to provide a function and inputs to be used for testing.

[sethaxen]

This is the best way I could think of to signal that Tracker should not be used for higher order AD.

[codecov-commenter]

Codecov Report

Merging #44 (6b7f2f1) into master (d9b3983) will increase coverage by 0.61%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #44      +/-   ##
==========================================
+ Coverage   82.39%   83.00%   +0.61%     
==========================================
  Files           4        5       +1     
  Lines         443      459      +16     
==========================================
+ Hits          365      381      +16     
  Misses         78       78              

Impacted Files	Coverage Δ
src/AbstractDifferentiation.jl	78.51% <100.00%> (+0.05%)	⬆️
src/tracker.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 d9b3983...6b7f2f1. Read the comment docs.

[sethaxen]

This PR is now ready for review.

Our version of jacobian is slower than Tracker's, but it seems to be less brittle. That being said, Tracker as a whole seems pretty brittle (e.g. Tracker.jacobian(cumsum, x) fails) and is likely not used much. Still, I think it should be included for now for completeness.

[mohamed82008]

Looks good!

[mohamed82008]

Thanks Seth and sorry for the delay.