slurm: fix session-stack lifetime and other plm/ras correctness bugs

[Petter-Programs]

Summary

Fixes #2546, plus a handful of related correctness issues found while
working through ras/slurm and plm/slurm.

• plm/slurm: the "using the whole allocation?" check compared
  map->num_new_daemons against session->nodes->size — a pointer
  array's capacity, not its populated count — so it emitted an
  explicit --nodes/--nodelist more often than intended.
• ras/slurm: prte_slurm_session_stack held a bare, unretained
  prte_session_t *, relying on the stack always popping an item
  before releasing its session. DVM finalize's blanket session release
  doesn't go through this stack, so a tracked session could be freed
  out from under it. The stack now retains its own reference on push
  and releases it via the stack item's destructor on every removal.
• ras/slurm: two spots (add_reused_nodes_to_session,
  add_pending_req) fed pmix_pointer_array_add's return value — an
  index, not a pmix_status_t — into prte_pmix_convert_status(),
  producing a meaningless error code on failure.
• ras/slurm: guard remove_nodes_by_count against ever computing
  zero nodes to remove in the "shrink and keep the job alive" branch;
  should be unreachable given the surrounding invariant, but the
  alternative (a NULL target list, or worse, an infinite shrink loop)
  is worse than failing loudly.
• ras/slurm: fixed a stale comment on kill_job's err_msg
  behavior.
• test/ras: updated the slurm module contract test for the
  session-stack change (below).
• examples/slurm: split the dvm command out of the reachability
  probe (previously run on every state query, so a probe failure was
  difficult to distinguish from an allocation failure) and dropped the
  README's srun wrapper around prterun, which is non-standard and
  could fail with all slots occupied.

release_allocation is now unimplemented everywhere

The session-stack retain change removes ras/slurm's
.release_allocation vtable entry — the reactive hook that decided
whether to scancel by checking if a session's item was still on the
stack. With the stack now holding a real reference, session_des can
never fire while an item is still linked, so the hook was structurally
dead; the scancel decision moves to being made explicitly at each
removal site instead.

ras/slurm was the only component that ever set
.release_allocation — no other ras component (hosts, pbs,
lsf, gridengine, flux, pmix, bootstrap, simulator,
testrm) implements it. Removing our one call site means the generic
plumbing around it — the release_allocation vtable slot itself,
prte_ras_base_release_allocation(), and its unconditional call from
session_des() on every session teardown, HNP and daemons alike — is
now dead code framework-wide. Left in place for this PR (only the
now-invalid unit test assertion was updated); flagging it as a
follow-up cleanup candidate rather than pulling it out here.

Testing

• Clean --enable-debug build, no warnings.
• make check (test/unit/ras/test_ras) passes.
• Session-stack retain/release change live-tested on a 3-node Slurm
  cluster (extend, shrink, and DVM shutdown mid-extension).

Credits: AccelCom @ Barcelona Supercomputing Center