BDDCML v2.2
major changes:

- added a self-contained example poisson_on_cube, solving Poisson problem on
cubic domain using cubic subdomains and cubic elements. Hopefully reasonably
commented to understand the data passed to BDDCML

- different default approach to identifying neighbour candidates, based on
overlaps of bounding boxes and overlaps of global indices.
More scalable than using parallel graph by ParMETIS, tested up to 65k cores

- changes to experimental interface scaling in BDDC

- fixed wrong C interface of version 2.1