Yet another forceeval

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

Add routine for TP flash for binary mixture, to solve for molar concentrations (and therefore mole fractions)

Found the first real bug in autodiff: https://github.com/autodiff/autodiff/issues/242

Not as fast as it could be, suspect the string conversion is not helping

Compilation of user code is very fast now

And add test

Could also be helped by perfect branch prediction