Return the number of prompt and generated tokens for API providers

[rlouf]

The number of prompt and generated tokens is important to evaluate the cost of each query. I suggest to return a class that contains the generation as well as these numbers.

[HerrIvan]

I suppose the goal is keeping the counts of the tokens consumed for each instance of the OpenAI class, right? Maybe keeping last and accumulated tokens, s.t. whoever is interested can inspect these values after each call.

[rlouf]

I was thinking about returning it at every step, but keeping it as a class attribute is much better!