fix: OpenAI Tool Return Serialisation

[hebenon]

Please describe the purpose of this pull request.
This PR is a potential fix for #3057. When a tool executes client-side (e.g., via letta-code) the resulting MessageRole.tool object persisted by the server contains populated tool_returns but an empty content list. The OpenAI Responses serializer in Message.to_openai_responses_dicts still assumes a single TextContent entry and asserts, causing every subsequent tool step to crash before the agent can continue.

Changes in the fix include:

• Reworked Message.to_openai_dict and Message.to_openai_responses_dicts so tool messages prefer the tool_returns payload (mirroring the Anthropic serializer) with fallback to legacy single-TextContent handling.
• Updated create_parallel_tool_messages_from_llm_response to inject a TextContent placeholder when only tool_returns are produced, preserving older consumers.
• Added test_message_serialization.py to validate both behaviors (tool-return-first and legacy fallback).

How to test
How can we test your PR during review? What commands should we run? What outcomes should we expect?

• Configure an agent to use an OpenAI model, with upserted tools from letta-cli.
• Give the agent a request that relies on local tools to complete.

Have you tested this PR?
Post-fix, tools can run in letta-cli with OpenAI agents. E.g.:

Related issues or PRs
This PR addresses #3057.

Is your PR over 500 lines of code?
No.

Additional context
For transparency, this fix was vibecoded with GitHub Copilot.