Zero pivot in ml_amg coarse solver

[Joachim Haga]

I sometimes see the PETSc-ML preconditioner failing with zero pivot in the coarse-grid solver.

The attached patch seems to help. I don't have the experience to tell if -mg_coarse_pc_factor_shift_type should be a parameter or if it can be set unconditionally like here.

Additionally, some minor fixes ("amg_ml" is wrong, plain "amg" can be either HYPRE or PCML, etc.).

[Joachim Haga (jobh)]

Oh, and perhaps PCML/PCHYPRE are undefined when PETSC_HAS_ML/PETSC_HAS_HYPRE are not set. If so, the patch is faulty (the mg_coarse_pc_factor_shift_type is valid, but not the other minor fixes).