This allows the user to set the required thread level. Initializing MPI
with threading support is needed when MPI functions are called from
multiple threads. All common MPI libraries seem to provide this feature.

Removed __cuda_callable__ from methods in DistributedArrayView, DistributedVectorView and DistributedMatrix

Distributed data structures are not supposed to be passed to device
kernels. Distributed data structures are operated by the host, which
uses the device for parallel processing only in the local data
structures.

It is useful especially when one wants to load mesh functions via the
readPointData or readCellData methods when the mesh was already loaded.

Also renamed PVTUReader's readLocalPointData and readLocalCellData to
fit into this hierarchy, hopefully it is clear that they return only
local data.

- ValueType must be non-const (otherwise it won't work with an array
  view with a const ValueType)
- IndexType should be taken from the array

Using references as class members is weird, because they cannot extend
the lifetime of the objects they are initialized with. Using values of
the type std::ostream and initializing by rdbuf (which is a pointer)
works better, probably because the underlying rdbuf generally outlives
the ostreams that were being passed to the writers.

Python cannot easily import modules containing "-" so we use "_dbg"
instead of "-dbg".

When the program is compiled without MPI support, or when run without
mpirun, the MpiCommunicator behaves just like NoDistrCommunicator.
Hence, we will not need to separate between the two classes.

To/matrices

See merge request !80

[skip ci]