DAX owns pex_config. CCing Fritz Mueller.

While wrapping meas_modelfit with pybind11 I was reminded how subtly fragile a lot of this stuff is; there are a lot of imports and some inspect calls to determine where a class should be imported from that just don't work very well, and the fact that these all tend to run at import time can make them very difficult to debug. That was all true under Swig, too, but the restructuring of Python modules we're doing in the pybind11 conversion upset the fragile balance in meas_modelfit and getting it working again was a pain. That package should probably be used to derived requirements for what the wrapping stuff should be able to do - the rest of the stack really doesn't exercise the hard cases (certainly the unit tests do not).

Another note: we should try to provide a C++ interface for creating Config objects from Control objects (which would probably have to import some pure-Python code, as the exceptions translation stuff does). Right now the need to declare config objects is a major reason we have to add pure-Python accompaniment to wrapper code in meas_modelfit (and the same is probably true in meas_base, though I haven't looked).

And another note: we should update the macros to include setting the default, since we can now do that in C++11.

Reviewed for DM-CCB, 2020-09-21. I'm not aware of any changes here.