Fixes to avoid data loss

[deruyter92]

Summary
Two defensive fixes on the GUI side to prevent silently losing processor
outputs when the DLC processor shuts down before the recording-stopped
hook fires.

The custom processor plugin dlc_inference_w_pd_sync saves its legacy
outputs (PROC pickle, DLC HDF5, timestamp NPY) and performs DB-compatible
file copies through on_recording_stopped().

some non-standard ways of ending the experiment could bypass
this hook

1. Any closeEvent: the shutdown skipped _cleanup_processor() when the worker thread
   stopped cleanly. After Fix thread crash in TeensyLatency + prevent data-loss for dlc_inference_w_pd_sync MMathisLab/FreelyMovingVR4Mice#329, the processor should now save it's own data when stopping. So cleanup of the processor is good here and safe.

2. Stop inference while recording: _stop_inference() stops
   the processor via _dlc.reset() before the recording stops. When recording is later stopped,
   _notify_processor_recording_stopped() finds no processor instance.

changes:

1. Add the missing self._cleanup_processor() call in the clean-shutdown
   branch of shutdown(), so that processor resources are properly released
   and (with the processor-side fix) data is saved before the window closes.
2. When the user clicks "Stop inference" while recording is active, a
   QMessageBox.question dialog now appears warning that legacy output
   copies will be skipped and recommending to stop recording first. If the
   user proceeds anyway, the processor saves its core data through its own
   stop() path. No prompt appears when recording is not active.

[C-Achard]

Few minor formatting comments, thanks for the fixes!

[C-Achard]

@deruyter92 These are all great fixes, I just need to rebase it on top of the current stack as the test/experimental fixes branch has diverged now, I marked it as draft but will take care of it now