Add benchmark tasks like BIG Bench

[kamalojasv181]

Whenever a new model is added, there need to be sufficient metrics that help the user estimate the drop in performance of the optimized model compared to the original model.

I suggest Big Bench. Other suggestions are welcomed.