Make I/O asynchronous per partition

[kordejong]

Current approach is:

• Group partitions per locality
• Iterate over localities:
  • Read or write the partitions
  • Wait for this to finish

This works fine for serial I/O. No need to hammer at the controller from multiple processes.

Drawbacks are:

• A partition that is read can participate in a computation, overlapping I/O of other partitions
• A ready partition can be written without waiting for other partitions, overlapping computations of other partitions
• Parallel I/O (independent) from multiple processes is fine