[0z5a]

Fixes #116.

Problem

nvshmemi_bootstrap() builds a bootstrap_attr_t whose initialize_shmem defaults to 0. It then
calls bootstrap_set_bootattr() with a NULL attribute when the application passed none, so the field
stays 0. With NVSHMEM_BOOTSTRAP=SHMEM and plain nvshmem_init(), the SHMEM plugin therefore skips
shmem_init(). It then calls shmem_my_pe() on an OpenSHMEM library that was never initialized.
#116 has the full trace.

Change

• bootstrap_set_bootattr() now sets initialize_shmem from flags & NVSHMEMX_INIT_WITH_SHMEM.
  Only an application that passes that flag has called shmem_init() itself.
• nvshmemi_bootstrap() always passes the attribute that bootstrap_init() uses.

shmem_finalize() already depends on nvshmem_initialized_shmem, so after this change the plugin
also finalizes what it initialized. The MPI plugin does the same with MPI_Init. The
NVSHMEMX_INIT_WITH_SHMEM path still passes 0 and is unchanged.

• When the plugin initializes OpenSHMEM itself, it keeps the OpenSHMEM library loaded:
  dladdr(shmem_init) finds the library, then dlopen(..., RTLD_NOLOAD | RTLD_NODELETE) pins it.
  In this case the application usually does not link OpenSHMEM, so unloading the plugin at
  nvshmem_finalize() also unloaded it. With Open MPI's OpenSHMEM, the process then crashed at exit
  (SIGSEGV) after shmem_finalize().

Testing

Setup: devel eaf955f vs this PR. Both columns use the host library and the SHMEM plugin built from
the same tree.

Open MPI 4.1 OpenSHMEM, 2 PEs on one node (B300). Launched with mpirun; same program binaries
on both sides.

case	devel	this PR
NVSHMEM_BOOTSTRAP=SHMEM, nvshmem_init(), nvshmem_malloc, barrier, nvshmem_finalize()	3/3: pshmem_my_pe: SHMEM is not initialized, abort	every run whose shmem_init() succeeded (3/3) passes on both PEs and exits 0
same, plugin without the library pin	—	3/3 pass on both PEs, then SIGSEGV at process exit (0/2 with LD_PRELOAD=liboshmem.so)
app calls shmem_init(), then NVSHMEMX_INIT_WITH_SHMEM	3/5 pass	2/5 pass

The remaining failures appear on both sides, including the unchanged NVSHMEMX_INIT_WITH_SHMEM
path. They are SHMEM_ABORT inside Open MPI's shmem_init() itself, after a memheap "too many
segments are registered" warning and before NVSHMEM prints anything. They are an Open MPI issue on
this host.

Sandia OpenSHMEM 1.5.3 (PMI-2), 1 PE (GH200).

case	devel	this PR
NVSHMEM_BOOTSTRAP=SHMEM, nvshmem_init(), nvshmem_malloc, barrier, nvshmem_finalize()	3/3 segfault (SOS: Out of symmetric memory, heap size 0)	3/3 pass, exit 0
app calls shmem_init(), then NVSHMEMX_INIT_WITH_SHMEM	3/3 pass	3/3 pass

An earlier revision without the library pin gave the same picture on SOS: 4/4 segfault on devel,
4/4 pass.

At 2 PEs, SOS 1.5.3 on that system fails its own first barrier (OFI error -22) without NVSHMEM
involved, so only 1 PE was possible there.