fix: duplicated reasoning tasks

[Palen]

• Prevent duplicated reasoning tasks by storing deep copies of newly added items in the streaming pending-items tracker, so _update_pending_items no longer mutates the responder’s original workflow item twice.
• Add regression test test_workflow_updates_not_applied_twice proving a responder that mutates a workflow before emitting updates yields only one reasoning task in both streamed and stored workflow items.

[jiwon-oai]

Thank you for this and apologies for the belated response! We merged this change in a separate PR 👍