Repartitioner for communication pattern

[greole]

This PR implements the first part of the repartitioner namely repartitioning of the communication pattern.

Repartitioning means that based on the parameter ranks_per_gpu a set of owning and non-owning ranks are created. Owning ranks take over all the matrix rows and communication of the non owning ranks.

To repartition the communication patter the following steps are performed:

• Check all communication partners on every rank and decide whether they are on a remote rank or local after repartitioning
• All communication to and from local after repartitioning ranks can be discarded
• The remaining non-local communication needs to be transferred to new owner rank

[chihta-wang]

It is a little confusing here.

According to the comment here, the function sorts a given set of interfaces into local and nonlocal interfaces.
But the function takes local and nonlocal sparsity patterns and returns a new non-local sparsity pattern.

Does the function actually do a sorting job for a given set of interfaces?

[greole]

Good catch, this comment is outdated. I'll fix it with the PR that implements the function.