- unused array mexx0 was allocated and never deallocated (noticed by
  Massimiaino Fatica)
- a temporary variable was not correctly initialized (most compilers set
  initial values to zero but this is not guaranteed)
- occupancies of Hubbard manifold (variable "ns") were computed independently
  on all processors, but their consistency was not ensured. Now there is a
  broadcast in "new_ns" that ensures consistency. While not a bug in principle,
  in practice this could lead to small divergencies buiding up between the
  ns on different processors, leading to small divergences of the
  self-consistency error (dr2), leading to small divergencies of the
  diagonalization threshold (ethr), leading to different numbers of
  unconverged eigenvalues on different processors, leading to mysterious MPI
  crashes. I am not sure about the origin of the divergence, but broadcasting
  ns seems to fix the problem.