Added a _selective_allreduce method to the slicing distributor in order to fix contractions that involve nodes with empty data.

-> Therefore modified the MPI_dummy.py

Fixed bincount method of _slicing_distributor (now uses allreduce instead of allgather+sum)

Added `copy` kwargs where missing.

Improved performance of binary operations in d2o since an unnecessary copy is now avoided.

Improved the d2o.cumsum method.

Extended the d2o docstrings.

Wrapped the demos code into if __name__ == "__main__" statements in order to make autodoc possible.

Made the steepest descent work.

propagator_operator still not working to full extend.

Fixed allreduce and Allreduce in MPI_dummy.py.
Added Option to d2o to skip the parsing of the input parameters. Added cumsum function to d2o.

Started working on the los response operator.
Switched off a sanity check in the d2o slicingdistributor's advanced indexing routine, because it was extremely slow (assert: is the first dimension ordered?).

Fixed zero-centering mask for even dimensions.

Some bugfixes.

Fixed a bug in nifty_mpi_data.