Fix a bug where retrying an export RQ job may break scheduling (#8584)

`_patched_retry` tries to schedule a copy of the current job. In
particular, it copies the dependencies of the old job using
`current_rq_job.dependency_ids`.

Unfortunately, `dependency_ids` does not return IDs of dependency jobs,
as one might expect. It actually returns the Redis _keys_ corresponding
to those jobs, as bytestrings. The RQ job creation code does not support
bytestrings as dependency specifiers, so it unintentionally treats them
as sequences, saving the individual bytes (as integers) as the
dependency job IDs. But since IDs have to be strings, the scheduler
quickly crashes when it tries to use those integer "IDs" to construct
Redis keys.

Thankfully, we don't actually need to get the dependency IDs.
`_patched_retry` is only used inside running jobs, and if a job is
running, it means that all its dependencies are already completed. Thus,
the newly scheduled job doesn't need to have any dependencies at all.

changelog.d/20241022_191618_roman_fix_integer_dependencies.md

@@ -0,0 +1,5 @@
+### Fixed
+
+- Fixed a bug where an export RQ job being retried may break scheduling
+  of new jobs
+  (<https://github.com/cvat-ai/cvat/pull/8584>)

cvat/apps/dataset_manager/views.py

@@ -85,7 +85,6 @@ def _patched_retry(*_1, **_2):
                 **current_rq_job.kwargs,
                 job_id=current_rq_job.id,
                 meta=current_rq_job.meta,
-                depends_on=current_rq_job.dependency_ids,
                 job_ttl=current_rq_job.ttl,
                 job_result_ttl=current_rq_job.result_ttl,
                 job_description=current_rq_job.description,