Asymmetric inplace operations

[niklas-uhl]

• gather
  • pass MPI_IN_PLACE on root, sendcount and sendtype ignore, contribution of root is assumed to be in the correct place in the buffer
• reduce
  • pass MPI_IN_PLACE as send_buf on root, data taken from recv_buf
• gatherv
  • pass MPI_IN_PLACE on root, sendcount and sendtype ignore, contribution of root is assumed to be in the correct place in the buffer
• scatter
  • pass MPI_IN_PLACE on root, sendcount and sendtype ignore, root just "keeps" vector to scatter
• scatterv
  • same as scatter
• alltoallv
  • scounts, displs, sendtype ignored
  • send and recv count between a processor pair must be the same -> limitation!

Originally posted by @niklas-uhl in #28 (comment)