fix(stirrup): restore tool messages for model calls

[syadav481]

Summary

• Restore NeMoUserMessage tool results back to ToolMessage before normal OpenAI-compatible chat-completions model calls.
• Preserve the existing agent-facing tool_response_as_user behavior while ensuring provider-facing payloads satisfy tool-call ordering rules.
• Preserve NeMoUserMessage tool results as function_call_output in exported run history and keep call IDs stable.

Why

#1181 fixed the summarization path, but normal Stirrup generation can still replay an assistant tool_calls message followed by a user-role NeMoUserMessage tool result into an OpenAI-compatible model server. Providers reject that payload because assistant tool calls must be followed by matching tool-role responses.

Testing

• ./.venv/bin/python -m pytest responses_api_agents/stirrup_agent/tests/test_app.py responses_api_agents/stirrup_agent/tests/test_nemo_client.py -q
• ./.venv/bin/ruff check responses_api_agents/stirrup_agent/nemo_agent.py responses_api_agents/stirrup_agent/nemo_client.py responses_api_agents/stirrup_agent/stirrup_utils.py responses_api_agents/stirrup_agent/tests/test_app.py responses_api_agents/stirrup_agent/tests/test_nemo_client.py

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.