Agent Factory `_Execute_model_async` Uses Non-Streaming `Ainvoke`, Preventing `On_llm_new_token` Callbacks

In `langchain/agents/factory.py`, the `_execute_model_async` method calls `await model_.ainvoke(messages)` instead of using streaming `astream`. Non-streaming invocation aggregates the full model response internally and never invokes the `on_llm_new_token` callback. LangGraph's `use_astream=True` streams graph events but does not change how the model node calls the LLM, so the model still returns a complete `AIMessage`.

1. Use langchain==1.3.4 and langgraph==1.2.4. 2. Create an agent with a callback handler that implements `on_llm_new_token`. 3. Run the agent with `astream_events` or graph streaming (e.g., `use_astream=True`). 4. Observe that the `on_llm_new_token` callback is never invoked, even though graph-level events are streamed.

Streaming aggregation may introduce overhead and memory usage proportional to response length. If the model does not support streaming or returns incomplete tool call chunks, the merge logic must handle `tool_call_chunks` to avoid losing tool calls. This change may alter timing and concurrency behavior; ensure `config` is not None when callbacks are required. Rollback: revert to the original `ainvoke` call if streaming causes regressions or tool call parsing issues.