Refactor prefill and inference benchmark

[IlyasMoutawwakil]

• fixes the issue of incompatible inputs between forward and generate passes by using generate for prefill.
• updates how we measure the processed volume (tokens/images/samples); taking into consideration cases where a text generation task takes other modalities as input.
• added an iterations argument to allow explicitly setting a minimum number of iterations a benchmark should perform; benchmark ends when both duration and iterations constraints are satisfied.