Add support for distributed I/O

[EiffL]

When eventually scaling up mesh computation to very large volumes, we need to output the layed out tensors to a file system. For cori and similar systems we will need to have each process output its own slice of the data. Not sure what's the best approach to do this yet.

[EiffL]

At least the parallel reading can be taken care of by this: https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/experimental/input_reader.py \o/