Force MUMPS ranks on a single shared memory node and fix for parmetis small graph size

[michel2323]

• The 4 MUMPS ranks are forced to be on a single shared memory node communicator that was created using MPI_COMM_TYPE_SHARED. See comments in the code for further details.
• Let MUMPS automatically choose whether to do parallel or sequential analysis
• Some verbosity fixes.