Enable with the cmake variable DUNE_CODEGEN_PROFILING. For each generated header,
a profile is written out. These can be visualized with tools like snakeviz.

CMakes add_custom_command occasionally (undeterministically)
trips over when multiple files are defined as output. The
result is unnecessary regeneration upon every change of the
accompanying C++ source...

The commit fixes the issue by defining one custom command per
operator. This has the additional pro of generating more verbose
messages during generation. Implementation wise, CMake needed
to extract more information from the ini file, so I switched
to using dune-testtools Python-CMake interface.

Instationary stuff not fixed so far.

So far, changes in the form compiler did not trigger code regeneration
if invoked from a downstream dune module. This is fixed now.

That allows us to load that file as a pandas frame.

based on dune-python!