fix(langchain): isolate middleware model calls from parent streaming …

[NITIN Madan (Nitin75408)]

Summary

• Add _get_internal_model_config() static method to AgentMiddleware
  base class that returns a RunnableConfig with callbacks=[] to
  prevent streaming callback inheritance from the parent agent context
• Update SummarizationMiddleware to use isolated config in both
  _create_summary() and _acreate_summary()
• Update LLMToolSelectorMiddleware to use isolated config in both
  wrap_model_call() and awrap_model_call() (previously passed no
  config at all)

Fixes #34382

Why

When astream() runs, var_child_runnable_config is set with streaming
callbacks. ensure_config() additively merges this into any config that
doesn't explicitly override callbacks. This caused internal middleware
model calls to emit tokens into the parent stream — breaking stream
correctness and isolation.

The fix passes callbacks=[] which ensure_config() merges on top of
the inherited config, clearing the streaming callbacks. This is backward
compatible: non-streaming usage already has no callbacks to clear.

Review notes

• SummarizationMiddleware previously passed config={"metadata": {...}}
  which preserved inherited callbacks. Now uses the isolated config.
• LLMToolSelectorMiddleware previously passed no config at all,
  fully inheriting parent callbacks. Now uses the isolated config.
• The _get_internal_model_config() method is intentionally private
  (_-prefixed) since it's an implementation detail, but available to
  custom middleware subclasses.