The 'aprun' job launcher can use protection domains.  These are a limited
resource and available only via aprun job launcher.  Other Cray systems, such
as cori, use srun, so we'll need this DRC mechanism to grant permission for
separate processes to talk to each other.

OpenMPi and MPICH both provide libmpi but vendors are slower to adopt this
change.

Confirm the open cache and stat cache are working.

Each provider will have to stat(2) a file once. In the expected case of
more clients than providers, we can save the additional system calls.

the 'bv_stat' client call has a side-effect of caching the information
locally, making the usual "broadcast from one" optimization infeasable.

Provider still calls older 'stat' system call but start laying the foundation
to use the more flexible statx

In some cases we need to re-open the file with the requested flags
(write and read).  Other sytem calls like flush and stat take an fd, but
we don't care how we got that fd.  no need to close and reopen in those
cases.

Guess I haven't built with "-Wall" in a while.

HDF5 (and its tests) makes use of MPI_File_set_size.  Benvolio parts in this
changeset.

We now create bootstrap-specific executables: e.g. you start benvolio
proveders with mpi using bv-server.mpi

return null and let caller deal with it

Provided some more documentation as well

The thallium 'expose' call boils down to a (posibly expensive) memory
registration call.

benvolio uses PMIx now.  Also fill in a few missing parts.

We are trying to remove build-time dependency on MPI.  'strided' and
'overlap' require mpi so we no longer build and execute them as part of
"make check".

updated 'wrap_runs' to mangage prte and pmix parts of job launching

bv_init now takes an object, not a config file, and relies on "something
else" (likely mpi) to efficiently distribute the state among clients.
To assist, a new bvutil_getconfig deals with converting the statefile
into such an object.