Split outstanding into active + outstanding; fix finalization race

[thatch]

The old single outstanding counter was overloaded: it covered both
batches with process() in flight and results sitting in
output_notifications waiting to be forwarded. That ambiguity let
_check_for_final see outstanding == 0 and mark a step final while
results were still queued, causing dropped notifications under
concurrency.

This change splits it into active (process() in flight) and
outstanding (items in output_notifications not yet forwarded).
_check_for_final now acquires state_lock once to read all three
counters as a consistent snapshot. _GuardedList and a __setattr__
override enforce the locking discipline statically.
unprocessed_notifications appends remain intentionally unlocked — they
only occur before the finalization window opens for a given step.

Also adds last_call(), a hook called exactly once when a step's inputs
are exhausted and all batches have drained, so subclasses can emit
trailing work without fighting the finalizer.