* This is configurable.

* Can be deconfigured for frequently changing client decompositions.

* Use distributed directory for intersection computation by default.

* Also add corresponding switch to test code.

* It is unclear if MPI I/O can make use of this, but it cannot hurt either.

* This version is still sub-optimal and transfers more data amongst the
  I/O server processes than strictly necessary.

* No component external to namespace needs to know.

* PIO client can only ever be invoked on active namespaces, other stages
  no longer exist.