v2.1.x: oshmem: move finalisation to oshmem_onexit

[ggouaillardet]

…nexit()

so we can use the legacy start_pes even when Open MPI is compiled with
--enable-static or --disable-visibility

(back-ported from commit 92dd719)

[jsquyres]

This is the v2.1.x version of #2120.

[ggouaillardet]

for the record, my initial analysis was incorrect.

i thought shmem_finalize() was never invoked under the hood if legacy start_pes() is used.
i ran under the debugger and found that the destructor is indeed always invoked, even if only static libraries are built.
what i missed is that opal_class_finalize(), which is also a destructor, is invoked before shmem_finalize(), which was also a destructor, so shmem_finalize() crashes when destroying objects.
/* this is true at least when i compile with --disable-shared --enable-static --disable-dlopen and in my environment */

i do not know if/how destructors can be ordered in the case of static libraries.

having shmem_finalize() invoked at exit (via on_exit()) ensures it is invoked before opal_class_finalize() so everything is fine.

[igor-ivanov]

@ggouaillardet see related comment at #2120 (comment)

[jsquyres]

We talked about this again on the call today. Everyone seems either neutral or ok with it: it solves a corner case that not many people will run into, but some will. So let's do it.

@ggouaillardet This needs to be rebased (perhaps because of the recent OSHMEM PRs?)

[hppritcha]

the travis failure is a known problem which has been fixed.