Now that we have tight Hub integration coming via #113, it could be useful to implement a simple benchmarking suite that allows users to:

• Select a dataset on the Hub
• Select a metric on the Hub
• Select N models (could already be optimised models)
• Optimise the models (if needed)
• Report a table of results comparing the gains in latency and impact on the model metric

In a first step, we might simply want to benchmark latency with some dummy input at various sequence lengths etc.