fix(openai-sdk): use AsyncSupermemory and correct client API calls

[icyzh]

The openai-sdk SupermemoryTools class initializes a sync Supermemory() client but all its methods are async and await it TypeError at runtime. On top of that, every .add() call targets the top-level client, but AsyncSupermemory exposes add() on the .memories sub-resource (client.memories.add(...)). Confirmed locally with the installed SDK:

`client.add(...)` raises `AttributeError: 'AsyncSupermemory' object has no attribute 'add'`
`client.memories.add(...)` works correctly
`client.memories` exists, `client` does not have `add`

That missing method means every memory-add call silently hits the except Exception handler and returns {"success": false, "error": "..."}. Even if it did reach the API, the raw Pydantic response gets fed to json.dumps() without .model_dump() another TypeError.

This rewrites the client init to use AsyncSupermemory, changes all .add() calls to .memories.add(), and calls .model_dump() on Pydantic responses before serialization. Test imports are also fixed ,the first fallback referenced a non-existent package name (supermemory_openai_sdk) and the second used a wrong module path.