Fix/persistency ingest lock

[viktorbeck98]

Task

#188

Description

EventPersistency.ingest_event() held no lock, so it could run concurrently with _save() / _load() on other threads:

• save vs ingest — the background _SaveTimer thread calls save() → _save() iterates ep.events_data while a concurrent ingest adds a new event_id → RuntimeError: dictionary changed size during iteration. This is swallowed by save()'s try/except, so the symptom is a silently-skipped save (logged as a warning), not a crash.
• load vs ingest — _load() rebinds events_data/event_templates/events_seen while ingest reads-then-writes them → torn/lost state. Currently prevented operationally by stopping the engine before /load (the service-api 409 guard).

Fix: give EventPersistency a single threading.RLock, acquire it in ingest_event(), and have PersistencySaver share the same lock so ingest / save / load are mutually exclusive.

• RLock (not Lock) is required: ingest fires the on_ingest callback → _check_event_count() → save() on the same thread, which re-acquires the lock. A plain Lock self-deadlocks here.
• PersistencySaver._lock is now aliased to persistency._lock, so existing save() / load() / locked() call sites are unchanged.

Out of scope (intentionally kept):

• The /load 409 guard stays. The lock makes load crash-safe, but load still wholesale-replaces live state — requiring the engine stopped is a semantics decision, not just a crash workaround.
• self._events_since_save += 1 remains unsynchronized (cosmetic save-cadence drift only).

Cost: one uncontended RLock acquire per event — negligible vs the existing msgpack/parquet work in add_data/dump.

How Has This Been Tested?

• Added TestPersistencySaverConcurrency::test_ingest_blocks_while_saver_lock_held (TDD): holds the saver lock on the main thread and asserts a concurrent ingest_event blocks until release. Fails on the old code (ingest_event ran while the saver lock was held), passes with the fix.
• The pre-existing test_events_until_save_triggers_save guards the RLock reentrancy regression — it deadlocks under a plain Lock, passes under RLock.
• Full suites green: 101 persistency tests + 109 detector tests (all of which drive ingest_event).
• prek passed locally (mypy, flake8, bandit, docformatter, etc.).

Checklist

• This Pull-Request goes to the development branch.
• I have successfully run prek locally.
• I have added tests to cover my changes.
• I have linked the issue-id to the task-description.
• I have performed a self-review of my own code.

[thorinaboenke]

can things can actually also get jumbled while data is exported?
import_state() uses saver and then does things with saver.locked() but export_state does plain persistency.save(),
should the export function also do

if saver is not None:
          with saver.locked():

?

[thorinaboenke]

same lock as persistency and event count timer use , so ingest_event now has to wait periodically for the state serialization (necessary) AND file writes. maybe worth moving the file write out of the critical section?