- Add finer caching control to selector for geometric operands.

- Add support of tensor variables (and tensor splitting when not
  allowed) in writers. Use _fvm_writer_tensor_order[] and
  fvm_writer_set_tensor_order to allow for different orderings).

- Add fvm_nodal_simplify_polygons() and fvm_nodal_simplify_polyhedra()
  functions.

- Add MPI communicator as argument to fvm_io_num_create()
  function (keeping track of communicator in structure),
  and replace use of fvm_parall_counter() type functions
  by similar mpi_io_num functions so as to allow for
  subcommunicators.
