Fixes#34139

When usingLLMToolSelectorMiddleware with stream_mode="messages", the internal tool selection tokens (the JSON output) are leaked into the final user stream. This happens because the internal LLM call inherits the parent's callback manager, which streams every token generated, even those intended only for middleware logic.

This PR introduces a fix to prevent the leakage:

1. Addeddisable_streaming parameter to the middleware init (defaults toTrue) to give users control.
2. Explicitly cleared Callbacks, instead of relying on model.bind(streaming=False) which is not respected by all model providers (e.g ChatOllama still leaks callback events) this implementation passes config={"callbacks": []} to the internal invoke and ainvoke calls.

I verified this locally using a reproduction script with ChatOllama (llama3.2).
Before: Output contained mixed internal JSON tokens({"tools": ...}).
After: Output is clean and contains only the final response tokens.