Extend external LammpsLibrary interface - #237

The previous commit tracked session ownership inside LibraryRunner
itself (_owns_lmp, checked in close(), with clear() issued at
construction time). Same outcome, wrong layer: LibraryRunner had to
know something that's really a property of the job, not the runner.

Move the decision to Phase.lammps_close, the method every driver stage
already calls when it's done with its runner for that stage:

    def lammps_close(self, lmp):
        if self.lmp is not None and self.calc.execution_mode == "library":
            lmp.clear()
        else:
            lmp.close()

self.lmp (set once in Phase.__init__, reused across every stage) is
exactly "was an external session ever supplied to this job" -- the
execution_mode check guards the one edge case where it wouldn't apply
(an lmp= passed with execution_mode="executable" is never touched by
create_object, so there is nothing to reuse). LibraryRunner goes back
to being a dumb wrapper: close() always closes, no ownership tracking.
Added clear() as its per-stage counterpart, called from lammps_close
instead of at construction -- the same effect, timed like the actual
end of a stage rather than the start of the next one.

Based on the same pattern from ICAMS#237 (jan-janssen,
"Extend external LammpsLibrary interface"), which solved this exact
problem the same way on the pre-refactor codebase before pylammpsmpi
support was removed and later re-added without carrying this over.

Tests reworked to match: dropped the _owns_lmp-specific ones, added
coverage for LibraryRunner.clear() directly and for lammps_close's
three branches (external+library -> clear, own session -> close,
external+non-library -> close), plus the same multi-stage reuse
regression test, now driven through job.lammps_close(...) instead of
calling LibraryRunner.close() directly.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>