Support run tokens

[andreasjansson]

"Run tokens" are used to run downstream Replicate models from within a model.

They are not yet exposed in an API, but can be found inside a Cog predictor with some ugly frame inspection:

def _find_api_token() -> str:
    frame = inspect.currentframe()
    while frame:
        if "self" in frame.f_locals:
            self = frame.f_locals["self"]
            if hasattr(self, "_current_run_token"):
                token = self._current_run_token
                return token
        frame = frame.f_back
    raise ValueError("No run token found in call stack")