Very fast in C++ (less than a microsecond for 4 outputs in the Jacobian), but 10 microseconds in Python for 2. Need to find a way to make it closer to original speed. Perhaps pass in a buffer to be populated.  That should allow the copy to be avoided. It must be the Eigen arrays that are so slow.

See #21

Add conversion routines for ideal-gas Helmholtz terms from CoolProp format

Yet another forceeval

Switch TP VLE solver to Powell's dogleg method. Might be just a little bit more reliable, but hard to tell.

See https://github.com/usnistgov/teqp/issues/9

More testing is still needed to identify other issues

Stopped iteration if dx values are all < 0

To check whether you have gotten to the critical point

* Split out reducing and mutants to their own headers

* Remove unused function

Based largely upon: https://pubs.acs.org/doi/abs/10.1021/acs.iecr.1c04703

TODO: apply polishing in another variable if constant composition polishing fails

Also not clear if this is the best way to include it.  Probably need the std::variant approach here too.

Python compilation working again

Break up pybind11 interface into separate compilation units (not as beneficial as hoped); still very slow