Feature Request: Token Usage Tracking (or model metadata?) Support

Related to #425, #263

Is there a consensus on implementing token usage tracking?

The most bare-bones option would probably be to just add a get_usage() method to BaseLanguageModel which returns the accumulated token usage of that model. Then that would be implemented for the supported models and a user could then call that method with a model from factory.create_model.

But maybe it'd be better to have a more general get_metadata() method for BaseLanguageModel that happens to return usage as part of it for the supported models.

For the general metadata option, maybe the output looks something like

{
  usage: {
    intput_tokens: int,
    output_tokens: int
  }
}

I'd be happy to contribute to this feature myself.