Fix Python 3.13 "Listener already started" in multiprocessing log listener

[AlecThomson]

Summary

• Fixes a Python 3.13-only RuntimeError: Listener already started raised on the second call to beamcon_3D.smooth_fits_cube (or beamcon_2D.smooth_fits_files) within the same process.
• racs_tools/logging.py builds a single module-level QueueListener at import time. Both functions called log_listener.start() on entry but only called enqueue_sentinel() on exit (never stop()), and every early-return/raise path (dry-run return, validation errors, "no files found") skipped even that. This left QueueListener._thread set after the monitor thread had already exited. Python 3.11/3.12 tolerate restarting in that state; Python 3.13's QueueListener.start() raises if _thread is not None.
• Adds a running_log_listener() context manager in racs_tools/logging.py that starts the listener and guarantees stop() runs in a finally block, then wraps the bodies of smooth_fits_cube and smooth_fits_files in it, removing the bare enqueue_sentinel() calls (calling both enqueue_sentinel() and stop() would double-sentinel the queue and silently break future runs — the fix removes the former so stop()'s own sentinel is the only one consumed).
• This unblocks running racs-tools under Python 3.13 — e.g. flint's flint_convol --mode convol --cubes, which calls smooth_fits_cube more than once per process (a dry-run get_cube_common_beam followed by convolve_cubes), and any long-lived worker (dask, celery) that outlives the task it ran.

Follow-up (not in this PR)

The listener is still a process-wide singleton, so two concurrent smooth_* calls in the same process would still interfere with each other. A fresh (logger, listener, queue) triple per call, with the queue passed explicitly into init_worker, would remove that class of bug entirely, but is a larger change than this fix warrants.

Test plan

• Added tests/test_logging.py:
  • test_smooth_fits_cube_twice: two dry runs of smooth_fits_cube in one process both succeed.
  • test_smooth_fits_cube_dryrun_then_real: a dry run followed by a real convolution (the order flint's CLI uses).
  • test_smooth_fits_cube_exception_then_success: an exception path (no files found) followed by a successful call.
  • test_running_log_listener_relays_records_across_calls: drives running_log_listener directly across repeated start/stop cycles and asserts worker log records still reach the handler each time (a stale sentinel from the old bug would silently break this).
  • Each test asserts log_listener._thread is None and the queue is empty after every call.
  • A guard_queue_listener_start fixture monkeypatches QueueListener.start to install Python 3.13's own guard, so these tests are meaningful (not a no-op) when run under 3.11/3.12.
• ruff check / ruff format --check pass on the changed files.
• Manually reproduced the original RuntimeError against the pre-fix code using the same guard, to confirm the regression tests actually exercise the bug.
• Note: could not run the full existing test suite (test_2d.py/test_3d.py) end-to-end in this environment since MIRIAD (fits/convol binaries) isn't installed here; those failures are pre-existing environment gaps unrelated to this change.

---

Generated by Claude Code