Benchmarking Suite

[curtischong]

I think it's time we have a script to benchmark TorchSim's models against each other and against ASE on various relaxation/MD tasks. The main reason why is because we can then make performance PRs more justifiable and reduce the burden of PR authors manually writing new performance testing scripts with each PR (note: I expect some PRs to include adhoc performance scripts - not all will be forgone).

At the very least we should be able to recreate the bar graph on the README with this script(s).

I don't think these scripts need to be run in CI, just run by PR authors when they submit a performance improvement.

[CompRhys]

What more is envisaged beyond https://github.com/TorchSim/torch-sim/blob/main/examples/scripts/8_bechmarking.py added by @falletta? This does get run in CI but as a smoke-test without assertions checking for regression

[curtischong]

I was thinking of benchmarking the functions on a real dataset (WBM?). It'd be nice if the script would produce the bar chart in the README.

[orionarcher]

I'd advocate for a script that can reproduce the figures in the TorchSim paper and can include any new models. We have potential GPU time from a couple external partners that we could leverage. It'd be nice to automatically run every time there is a new release and store the results somewhere in the docs.