ENH: Add metastable composition sets to starting point

[richardotis]

This PR adds a new function, add_nearly_stable, which augments the starting point of the solver with additional CompositionSet objects, which are entered with zero amount. These objects are added by the following criteria:

• The phase does not already exist on the starting hyperplane (this will not add missing composition sets of a miscibility gap)
• For a given phase, the added configuration is the closest to the starting hyperplane, within 1000 J/mol-atom

The purpose of this change is to improve the accuracy of solutions when sampling densities are not high enough to find the true set of equilibrium composition sets. This can easily happen in the multi-component / multi-sublattice case.

[codecov]

Codecov Report

Merging #362 (7043f23) into develop (e9c5d6d) will increase coverage by 0.04%.
The diff coverage is 100.00%.

@@             Coverage Diff             @@
##           develop     #362      +/-   ##
===========================================
+ Coverage    89.86%   89.91%   +0.04%     
===========================================
  Files           44       44              
  Lines         4341     4362      +21     
===========================================
+ Hits          3901     3922      +21     
  Misses         440      440              

Impacted Files	Coverage Δ
pycalphad/core/calculate.py	92.63% <100.00%> (+0.15%)	⬆️
pycalphad/tests/datasets.py	100.00% <100.00%> (ø)
pycalphad/tests/test_equilibrium.py	97.65% <100.00%> (+0.08%)	⬆️
pycalphad/tests/test_database.py	100.00% <0.00%> (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 e9c5d6d...7043f23. Read the comment docs.

[bocklund]

Does this need a test and what does that look like? Do we need a unit test that checks that composition sets are added or is an integration test for a tricky case enough?

[richardotis]

Does this need a test and what does that look like? Do we need a unit test that checks that composition sets are added or is an integration test for a tricky case enough?

I think an integration test for a test case is the right idea, because the use of metastable composition sets in the starting point is an implementation detail. Do you have a good test case? I hesitate to use the multi-component system that inspired this because of the additional test time.

[bocklund]

Thanks for putting together this PR!