- Require only mpi4py (all MPI operations done in python).  This
   allows us to build our compiled extensions with any compiler
   rather than requiring consistency with the compilers used to
   build mpi4py.

 - Port the high-level C++ atmosphere simulation code to python
   and use compiled code only on a single process.  Change the
   on-disk format of cached atmosphere realizations to use hdf5.

 - Update build system to not look for MPI compilers.

 - Add a new log level, "VERBOSE" which is even lower than "DEBUG".
   The intention for this is to contain messages that might come
   from every single process.

 - The MPIShared and MPILock classes are found in the upstream
   pshmem package.  Bundle a tagged release of that package for
   now until it is widely available on PyPI / conda-forge.

 - Clean up the generation of version.cpp so that it only happens
   if the version actually changes.

…nto place at the end.

function is actually called.  TOAST does not use numba since it conflicts
with our existing explicit parallelism.  The set_numba_threading() function
is designed to be used before calling external packages that use numba,
in order to ensure consistency between toast and numba.

… that the kolmo_x spectrum was not being reduced across processes.