Allreduce custom operations

[Tehforsch]

There are two paths to go here:

1. Implement all the basic operations explicitly, using the SystemOperations of rsmpi and again for the local communication. This will allow disabling the user_operations feature in rsmpi which should remove the dependency on libffi.
2. Allow user operations. rsmpi makes this easy, but implementing them for the local communication will be tricky, since the closures for rsmpi depend on DynBuffer types, which will probably be hard to construct without access to MPI. The way to go here might be to write a wrapper that can be called conveniently with a closure which will then be translated into the appropriate closure in MPI runs and just simply executed in local runs.