ch4/{ofi,ucx}: Skip collective finalize in sessions model

If there is no MPI_COMM_WORLD, we should skip calling a global PMI
barrier during finalization. It is not guaranteed that world processes
have initialized MPI (thus PMI), so a barrier could hang.

Author: raffenet

src/mpid/ch4/netmod/ofi/ofi_init.c

@@ -925,7 +925,7 @@ int MPIDI_OFI_mpi_finalize_hook(void)
     /* Destroy RMA key allocator */
     MPIDI_OFI_mr_key_allocator_destroy();
 
-    if (MPIR_CVAR_NO_COLLECTIVE_FINALIZE) {
+    if (!MPIR_Process.comm_world || MPIR_CVAR_NO_COLLECTIVE_FINALIZE) {
         /* skip collective work arounds */
     } else if (strcmp("verbs;ofi_rxm", MPIDI_OFI_global.prov_use[0]->fabric_attr->prov_name) == 0
                || strcmp("psm2", MPIDI_OFI_global.prov_use[0]->fabric_attr->prov_name) == 0

src/mpid/ch4/netmod/ucx/ucx_init.c

@@ -243,7 +243,7 @@ int MPIDI_UCX_mpi_finalize_hook(void)
     int n = MPIDI_UCX_global.num_vcis;
     pending = MPL_malloc(sizeof(ucs_status_ptr_t) * MPIR_Process.size * n * n, MPL_MEM_OTHER);
 
-    if (!MPIR_CVAR_NO_COLLECTIVE_FINALIZE) {
+    if (MPIR_Process.comm_world && !MPIR_CVAR_NO_COLLECTIVE_FINALIZE) {
         /* if some process are not present, the disconnect may timeout and give errors */
         mpi_errno = MPIR_pmi_barrier();
         MPIR_ERR_CHECK(mpi_errno);