Don't call the delete callback on MPI_COMM_WORLD attributes.

[bosilca]

Fixes #12035.

[jsquyres]

For an MPI app that caches no attributes on MPI_COMM_WORLD, deleting this code doesn't orphan or leak any memory during finalization, right?

[bosilca]

Correct. Any attributes attached to MPI_COMM_WORLD once MPI_Finalize is called will leak.

[jsquyres]

Correct. Any attributes attached to MPI_COMM_WORLD once MPI_Finalize is called will leak.

That's not quite what I'm asking.

I notice that you're deleting an OBJ_RELEASE(ompi_mpi_comm_world.comm.c_keyhash). Is that memory still freed properly during the OBJ_DESTRUCT() of COMM_WORLD? E.g., did you test running with and without this PR to verify that the amount of orphaned / leaked memory is the same? (I'm guessing that we're not overall Valgrind clean, so I'm just asking to make sure that we're not creating a new leak by deleting this code)

[bosilca]

Thanks @jsquyres. Based on your comment I realized that the code I remove was in fact partially useful, as it had the side effect of removing the predefined attributes we were attaching to MPI_COMM_WORLD. Unfortunately, it had the side effect of also deleting user-defined attributes, leading to the issue #12035. Please recheck.