xyst

Replace graph aggregation with scalable algorithm

Using a graph-based partitioner with zoltan requires a unique graph.
Since the mesh is read in a distributed fashion, this requires
communication. So far this was done simply aggregating the full graph on
each compute node which does not scale well, evidenced by out of memory
during this step even with moderate-size, e.g., nelem=O(1e+6),  meshes.

This commit replaces the graph aggregation, preceding the call to
zoltan's phg, by constructing a distributed table, storing the
aggregated graph, assembled by a query-response algorithm. This
algorithm assigns compute nodes as owners whose task is to aggregate
points surrounding points (the graph for a given global node id). If
multiple compute nodes contribute to a single global mesh node, the
compute node with a lower id is the owner.

Compared to the previous one, this algorithm requires only
point-to-point communication among a limited set of compute nodes, that
share part of the graph adjacent to the global id owned. As a result,
this scales much, better than the previous one, using less memory.

The result changes slightly with a chance of increasing randomness in
the final result (as this data is handed to zoltan for partitioning).
This required some minor adjustments to a couple of regression tests.