Refactor of reverse transfers and fixes for FD group that contains parallel group(s).

[naylor-b]

Summary

• Refactored reverse transfers so that internal derivative values are correct throughout the model instead of being off by some multiple, determined by the number of duplicated responses, which we previously corrected for at the end of the total derivative calculation.
• Fixed an issue in both fwd and rev modes where total derivatives were incorrect in cases where a group using FD or CS contained a parallel group.
• Added a command line tool to display the full set of connections across ranks when running under MPI.
• Fixed a number of bugs that resulted from the way we were doing reverse transfers previously.

Related Issues

• Resolves Can't compute derivatives for a Group using finite difference if it contains a ParallelGroup #3046
• Resolves Parallel derivative coloring with vector design variable can yield zero derivatives #2919
• Resolves Incorrect derivatives when using bcasts and parallel groups connected to outer components #2883

Backwards incompatibilities

This could break models where someone has added code to correct for the previously incorrect derivatives by dividing by the comm size or something similar.

New Dependencies

None

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[coveralls]

coverage: 89.75% (-0.3%) from 90.021%
when pulling 6e16d7c on naylor-b:transfers5
into fee4510 on OpenMDAO:master.