1) Add a frotran library test case (sequential + automatic io)

2) Improve the existing C test cases (clean deallocations in case of errors)
3) Remove a bug in PMMG_DEL_MEM Macro (the memory counter was decremented even if the macro was called on a null pointer)
4) Call PMMG* memory macros in merge_mesh
5) Remove various errors in the memory counter leading to non zero values at end
6) Try to have MPI safe merge_parmesh + check_ext_comm* functions even when it fails
7) Begin to handle the case when parmetis give an empty partition. It doesn't work in fact because parmetis doesn't accept an empty local graph as input.