fix(tools): support multi-part and array message content in OpenAI middleware

[Adityakk9031]

Problem

In packages/tools/src/openai/middleware.ts, getLastUserMessage and getConversationContent previously assumed message content was strictly a string.

When callers use modern OpenAI multimodal or multi-part messages where content is an array of content parts ([{ type: "text", text: "..." }, ...]), typeof content === "string" evaluated to false, causing getLastUserMessage to return "".

In createWithMemory, !userMessage evaluated to true, which caused the middleware to:

1. Silently skip semantic memory retrieval.
2. Skip saving the conversation context to memory.
3. Pass through the completion request without injecting any Supermemory context.

Solution

1. Added extractMessageText in openai/middleware.ts to cleanly extract and join text parts from both string content and array content parts (ChatCompletionContentPartText).
2. Updated getLastUserMessage and getConversationContent to use extractMessageText.
3. Added unit tests under packages/tools/src/openai-middleware.test.ts covering string, multi-part, and multimodal inputs.

Verification

Ran bun test packages/tools/src/openai-middleware.test.ts (all 6 tests passed).

[Adityakk9031]

Hey @Dhravya @MaheshtheDev, could you please take a look at this fix when you have a chance? Thank you!