No longer attempting to use MPI_COMM_WORLD as default anywhere in library code

[edhartnett]

Fixes #989.

We can, should, and do use MPI_COMM_WORLD in our tests, but not in the library code. We must only operate on the communicator specified in the open/create, otherwise we risk hanging the users MPI code.

[WardF]

Going to play catchup and aggregate a few including this one and the hdf5.h one today.

[edhartnett]

Thanks!

Actually this PR already includes the hdf5.h fix, as do a bunch of others, since I need it to pass my own CI system. So you can just merge one of the others and I will close the hdf5.h one. ;-)

#980 is the one I am waiting for. The rest are just me closing tickets while I wait for #980. ;-)

#967 should not be merged until after Dennis' #558, to avoid stepping on our toes with some of the generated files that are checked into the repo. Once #558 is merged, I need to update #967 before it gets merged.

[edhartnett]

@WardF also note that Dennis' current inmemory PR is merged into a bunch of my open PRs, in order to pass CI. This makes the code review confusing, since it looks like I messed about with the diskess settings in configure and makefiles. Once Dennis' PR is merged, many of mine will get simpler, so perhaps you should start with his.

[edhartnett]

@WardF one of the travis tests timed out trying to load. But these changes pass my full CI and are ready to merge.

[edhartnett]

OK, all travis builds passed. This PR is ready for merging.

[WardF]

Code review of this and the other open @edhartnett PR is in progress now.