fix: avoid replaying transient Responses API items through PatchedChatOpenAI

[supper-ming]

Summary

This PR makes PatchedChatOpenAI safe to use with OpenAI-compatible Responses API gateways that do not persist response items when store=false.
Changes:

• keep the existing Gemini thought_signature replay fix
• strip non-replayable Responses API items such as reasoning blocks from replayed
  history when store is not enabled
• preserve replayable tool-call items like function_call and function_call_output
• add regression tests for both store=false and store=true
• clarify the configuration docs for OpenAI-compatible Responses API gateways

Root Cause

In multi-turn conversations, DeerFlow replays prior assistant history back into the next
Responses API request.

For some OpenAI-compatible gateways, when store=false, response-only items such as
reasoning are not persisted server-side. Replaying those opaque items in the next turn
can fail with errors like:

Item with id 'rs_...' not found. Items are not persisted when `store` is set to false.

This surfaced in DeerFlow after clarification/tool-driven turns, where LangChain forwarded
prior Responses API items back into input unchanged.

## Why this fix

PatchedChatOpenAI already exists to handle OpenAI-compatible gateway quirks, specifically
Gemini thought_signature replay. Extending it to also sanitize non-replayable Responses
API items keeps the fix localized to the compatibility layer and avoids changing normal
thread or frontend behavior.

The patch:

- removes transient response items only when store is not explicitly enabled
- keeps tool-call continuity intact
- preserves existing behavior when store=true

## Tests

Ran:

uv run --project backend --group dev python -m pytest backend/tests/test_patched_openai.py
-q

Result:

10 passed

## Notes

This PR intentionally does not include:

- local environment/port changes
- frontend thread/navigation changes

It is scoped only to the confirmed backend compatibility issue.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}