- The --version argument is now passed down to the "real" mpi wrapper,
  which is then passed on to the serial compiler.
- The mpif77 and mpif90 wrappers are removed, and replaced with the
  mpifort wrapper.
- mpiCC and mpic++ are installed as symlinks to mpicxx
- mpif77 and mpif90 are installed as symlinks to mpifort

- Add missing constant MPI_MAX_LIBRARY_VERSION_STRING
- Add missing constant MPI_ERRCODES_IGNORE
- Add missing function MPI_Comm_spawn

- Added "YogiMPI_VERSION_STR" to yogimpi.h and mpif.h, which should be
  set to the version of YogiMPI used at compile time.
- Added "make dist" target, which will create a yogimpi-<VERSION>.tar.gz
  file, where <VERSION> is determined by $(git describe --tags).

By inspection, there is more to do here.

Before this commit, when client code would pass a copy or delete
callback to MPI_Comm_create_keyval, if the callbacks had any MPI calls
inside them then the code would segfault.  The segfault was due to the
native MPI passing in the native MPI_Comm to the callback, and then
YogiMPI would segfault when tyring to covert the comm to a MPI_Comm
because Yogi did not know the comm was already converted.

This change captures the callback functions and wraps them in lambdas
that un-converts the comm back to a YogiMPI_Comm so it is safe for the
callbacks to contain MPI calls.

There are likely other failure modes associated with user-supplied
callbacks to MPI routines.  Hopefully this will be a good pattern to
repeat if we find similar issues with other callbacks.

This commit wraps almost every single native MPI call with a callDepth
increment/decrement.  Currently this is only used for indenting calls in
the yogimpi.log.* files.

There are a certain class of problems that are hard to diagnose, and the
commonality is that YogiMPI has called the native MPI and then that MPI
has called a callback that was wrapped by YogiMPI.  Adding this
callDepth and indenting the log file based on callDepth makes it much
easier to diagnose the problem.

This commit wraps MPI_User_functions in lambdas, so we can revert the
MPI_Datatype back to a YogiMPI_Datatype before passing it to the
callback so that the callback can be called from the native MPI routines
(e.g. MPI_Reduce, MPI_Scan, etc.).

This reverts commit de7576f.

This reverts commit a9bf7ae.

This reverts commit cc8ec8d.

This reverts commit 53c3c0c.

Still have more testing to do and a bit more cleanup, but I need to move
my exploration to another machine.

Now only datatypes are using the new insertIntoPoolIfNotExists call.

When using the new call for comms and requests things stop working for
some reason I don't yet understand, so I'm only using the new call for
the datatype which was causing the most issues (I think ...).