Add support for parallel IO with mpi4py

[jswhit]

add parallel IO capabilities. netcdf-c and hdf5 must be compiled with MPI support, and mpi4py must be installed. To open a file for parallel access, use parallel=True in Dataset.__init__ and optionally pass the mpi4py Comm instance using the comm kwarg and the mpi4py Info instance using the info kwarg. If neither are specified, MPI_COMM_WORLD and MPI_INFO_NULL are used. IO can be toggled between collective and independent using Variable.set_collective. See examples/mpi_example.py and pull request #716.

[jswhit]

There are no tests yet, but for some reason tst_diskless.py fails when the an mpi-enabled netcdf-c library is linked.

[jswhit]

Travis now runs examples/mpi_example.py. Failures are allowed for now while this support is experimental.

[jswhit]

Seems ready to go, will merge tomorrow if there are no comments.

[jswhit]

pull request #716 merged

[ArchangeGabriel]

@jswhit Where you able to pin/fix the tst_diskless.py failure? I had reported it a while ago #669 so that some tests had already been done around it.

[jswhit]

No - someone should probably create at netcdf-c issue for this.