Reimplement emcee_hammer

New implementation is much more compact and hopefully simple to understand.
Plus, it utilizes direct calls to the fit function instead of using eval_counts.
This has the benefit of performance (at least the documentation claims this)
and that the model evaluation takes the free parameters directly (so no calls
to get_par or equiv. required).

Currently the implementation uses the rcl-mpi module, which is not a clean
implementation of the mpi routines. A future version should fix this if possible