Issue encountered

I need to evaluate google translate and other methods that do not fit in any of the evaluation modes supported currently.

Solution/Feature

I am not sure what the best way to do this is, but I see the following general solution:
We add one ModelConfig, where the user can specify a function that is called for a batch of examples. As long as the user provides the output correctly formatted, they can evaluate any system they like. It does not need to fit into the model_configs supported so far.
Happy to propose a PR for this.