queue_job: job_record_with_same_identity_key ignore STARTED

Good explanation from Guewen:

Depending on the use case, we should or should not include.
As part of my current work, I use Sidekiq daily.
They have a similar feature, but you can set a per-job parameter unique_until with options:
start (that would mean up to ENQUEUED here) or success (that would be up to STARTED here).

Think about this use case: a job refreshes a cache. Data have changed, we create a pending job. Data change again, no new job because the job is still pending.
Job starts. Data change while the job is running. In this very case, we'd like to enqueue a new job otherwise the cache will be outdated.

I reckon that both cases are valid, but I fear adding this state in the domain may, silently and in subtle ways, existing behaviors.

queue_job/job.py

@@ -300,7 +300,7 @@ def job_record_with_same_identity_key(self):
             .search(
                 [
                     ("identity_key", "=", self.identity_key),
-                    ("state", "in", [WAIT_DEPENDENCIES, PENDING, ENQUEUED, STARTED]),
+                    ("state", "in", [WAIT_DEPENDENCIES, PENDING, ENQUEUED]),
                 ],
                 limit=1,
             )