feat(langchain): model call limits (#33178)
This PR adds a model call limit middleware that helps to manage:
* number of model calls during a run (helps w/ avoiding tool calling
loops) - implemented w/ `UntrackedValue`
* number of model calls on a thread (helps w/ avoiding lengthy convos) -
standard state
Concern here is w/ other middlewares overwriting the model call count...
we could use a `_` prefixed field?
